rm extract

.history/app_20250404171116.py

@@ -0,0 +1,329 @@
+import gradio as gr
+from pathlib import Path
+from reactagent.environment import Environment
+from reactagent.agents.agent_research import ResearchAgent
+from reactagent.runner import create_parser
+from reactagent import llm
+from reactagent.users.user import User
+import os
+import json
+
+
+# Global variables to store session state
+env = None
+agent = None
+state_example = False
+state_extract = False
+state_generate = False
+state_agent = False
+state_complete = False
+index_ex = "1"
+
+example_text = [
+    "Research Paper 1: Dataset and Baseline for Automatic Student Feedback Analysis",
+    "Research Paper 2: An Empirical Study on the Impact of Code Review on Software Quality"
+]
+
+# Load example JSON file
+def load_example_data():
+    with open("example/example_data.json", "r") as json_file:
+        example_data = json.load(json_file)
+    
+    for idx in example_data.keys():
+        try:
+            file = example_data[idx]["code_init"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_init"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")
+        try:
+            file = example_data[idx]["code_final"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_final"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")     
+    return example_data
+
+example_data = load_example_data()
+
+# Function to handle the selection of an example and populate the respective fields
+def load_example(example_id):
+    global index_ex 
+    index_ex = str(example_id)
+    example = example_data[index_ex]
+    paper_text = 'Title:\t' +  example['title'] + '\n\nAbstract:\t' + example['abstract']
+    return paper_text
+
+example_text = [load_example(1), load_example(2)]
+
+# Function to handle example clicks
+def load_example_and_set_index(paper_text_input):
+    global index_ex, state_example
+    state_example = True
+    index_ex = str(example_text.index(paper_text_input) + 1)
+    paper_text =  load_example(index_ex)
+
+    return paper_text,  "", "", "", "", "", ""
+
+
+
+########## Phase 1 ##############
+
+def extract_research_elements(paper_text):
+    global state_extract, index_ex, state_example
+    if not state_example or paper_text == "":
+        return "", "", "", ""
+    state_extract = True
+    if not paper_text.strip().startswith("Title:\t" + example_data[index_ex]["title"]):
+        print("Mismatch detected.")
+        print(paper_text)
+        return "", "", "", ""
+    example = example_data[index_ex]
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return tasks, gaps, keywords, recent_works
+
+
+# Step 2: Generate Research Hypothesis and Experiment Plan
+def generate_and_store(paper_text, tasks, gaps, keywords, recent_works):
+    if (not state_extract or not state_example or paper_text == ""):
+        return "", "", "", ""
+    global state_generate, index_ex
+    state_generate = True
+    hypothesis = example_data[index_ex]['hypothesis']
+    experiment_plan = example_data[index_ex]['experiment_plan']
+    return hypothesis, experiment_plan, hypothesis, experiment_plan
+
+##########  Phase 2 & 3  ##############
+def start_experiment_agent(hypothesis, plan):
+    if (not state_extract or not state_generate or not state_example):
+        return "", "", ""
+    global state_agent, step_index, state_complete
+    state_agent = True
+    step_index = 0
+    state_complete = False
+    # predefined_message = f"Implement the following hypothesis and experiment plan:\n\nHypothesis:\n{hypothesis}\n\nExperiment Plan:\n{plan}"
+    return example_data[index_ex]['code_init'], predefined_action_log, "", ""
+
+def submit_feedback(user_feedback, history, previous_response):
+    if (not state_extract or not  state_generate or not state_agent or not state_example):
+        return "", "", ""
+    global step_index, state_complete
+    step_index += 1
+    msg = history
+    if step_index < len(process_steps):
+        msg += previous_response + "\nUser feedback:" + user_feedback + "\n\n"
+        response_info = process_steps[step_index]
+        response = info_to_message(response_info)  # Convert dictionary to formatted string
+        response += "Please provide feedback based on the history, response entries, and observation, and questions: "
+        step_index += 1
+        msg += response
+    else:
+        state_complete = True
+        response = "Agent Finished."
+        
+    return msg, response, example_data[index_ex]['code_init'] if state_complete else example_data[index_ex]['code_final'], ""
+
+def load_phase_2_inputs(hypothesis, plan):
+    return hypothesis, plan, "# Code implementation will be displayed here after Start ExperimentAgent."
+
+
+
+predefined_action_log = """
+[Reasoning]: To understand the initial structure and functionality of train.py for effective improvements.
+[Action]: Inspect Script (train.py)
+Input: {"script_name": "train.py", "start_line_number": "1", "end_line_number": "74"}
+Objective: Understand the training script, including data processing, [...]
+[Observation]: The train.py script imports [...]. Sets random seeds [...]. Defines [...] Placeholder functions [...] exist without implementation. [...]
+[Feedback]: The script structure is clear, but key functions (train_model, predict) need proper implementation for proposed model training and prediction.\n
+"""
+
+
+predefined_observation = """
+Epoch [1/10],
+Train MSE: 0.543,
+Test MSE: 0.688
+Epoch [2/10],
+Train MSE: 0.242,
+Test MSE: 0.493\n
+"""
+
+# Initialize the global step_index and history
+process_steps = [
+    {
+        "Action": "Inspect Script Lines (train.py)",
+        "Observation": (
+            "The train.py script imports necessary libraries (e.g., pandas, sklearn, torch). "
+            "Sets random seeds for reproducibility. Defines compute_metrics_for_regression function "
+            "to calculate RMSE for different dimensions. Placeholder functions train_model and "
+            "predict exist without implementations."
+        ),
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The script executed successfully. Generated embeddings using the BERT model. Completed "
+            "the training process without errors. Metrics calculation placeholders indicated areas needing implementation."
+        ),
+    },
+    {
+        "Action": "Edit Script (train.py)",
+        "Observation": (
+            "Edited train.py to separate data loading, model definition, training loop, and evaluation into distinct functions. "
+            "The edited train.py now has clearly defined functions"
+            "for data loading (load_data), model definition (build_model), "
+            "training (train_model), and evaluation (evaluate_model). Similarly, eval.py is reorganized to load the model and perform predictions efficiently."
+        ),
+    },
+    {
+        "Action": "Retrieve Model",
+        "Observation": "CNN and BiLSTM retrieved.",
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The model trained over the specified number of epochs. Training and validation loss values are recorded for each epoch, "
+            "the decrease in loss indicates improved model performance."
+        )
+    },
+    {
+        "Action": "Evaluation",
+        "Observation": predefined_observation,
+    }
+]
+def info_to_message(info):
+    msg = ""
+    for k, v in info.items():
+        if isinstance(v, dict):
+            tempv = v
+            v = ""
+            for k2, v2 in tempv.items():
+                v += f"{k2}:\n  {v2}\n"
+        v = User.indent_text(v, 2)
+        msg += '-' * 64
+        msg += '\n'
+        msg += f"{k}:\n{v}\n"
+    return msg
+
+
+def handle_example_click(example_index):
+    global index_ex
+    index_ex = example_index
+    return load_example(index_ex)  # Simply return the text to display it in the textbox
+
+# Gradio Interface
+with gr.Blocks(css=".gr-examples-label {display: none;}", theme=gr.themes.Default()) as app:
+    gr.Markdown("# MLR- Copilot: Machine Learning Research based on LLM Agents")
+    gr.Markdown("###   ")
+    gr.Markdown("## <span style='color:Orange;'> This UI is for predefined example demo only.</span>")
+    gr.Markdown("## <span style='color:Orange;'> To reproduce the results please use [Github](https://github.com/du-nlp-lab/MLR-Copilot/)</span>")
+
+
+
+    gr.Markdown("MLR-Copilot is a framework where LLMs mimic researchers’ thought processes, designed to enhance the productivity of machine learning research by automating the generation and implementation of research ideas. It begins with a research paper, autonomously generating and validating these ideas, while incorporating human feedback to help reach executable research outcomes.")
+
+
+    gr.Markdown("## ⬇️  Click an example at bottom to start·！")
+
+
+    # Use state variables to store generated hypothesis and experiment plan
+    hypothesis_state = gr.State("")
+    experiment_plan_state = gr.State("")
+
+    ########## Phase 1:  Research Idea Generation Tab ##############
+    with gr.Tab("💡Stage 1: Research Idea Generation"):
+        gr.Markdown("### Extract Research Elements and Generate Research Ideas")
+        with gr.Row():
+            with gr.Column():
+                paper_text_input = gr.Textbox(value="", lines=10, label="📑 Research Paper Text", interactive=False)
+
+                extract_button = gr.Button("🔍 Extract Research Elements")
+                with gr.Row():
+                    tasks_output = gr.Textbox(placeholder="Research task definition", label="Research Tasks", lines=2, interactive=False)
+                    gaps_output = gr.Textbox(placeholder="Research gaps of current works", label="Research Gaps", lines=2, interactive=False)
+                    keywords_output = gr.Textbox(placeholder="Paper keywords", label="Keywords", lines=2, interactive=False)
+                    recent_works_output = gr.Textbox(placeholder="Recent works extracted from Semantic Scholar", label="Recent Works", lines=2, interactive=False)
+            with gr.Column():
+                with gr.Row():  # Move the button to the top 
+                    generate_button = gr.Button("✍️ Generate Research Hypothesis & Experiment Plan")
+                with gr.Group():
+                    gr.Markdown("### 🌟 Research Idea")
+                    with gr.Row():
+                        hypothesis_output = gr.Textbox(label="Generated Hypothesis", lines=20, interactive=False)
+                        experiment_plan_output = gr.Textbox(label="Generated Experiment Plan", lines=20, interactive=False)
+        
+    
+        # Step 1: Extract Research Elements
+        extract_button.click(
+            fn=extract_research_elements, 
+            inputs=paper_text_input, 
+            outputs=[tasks_output, gaps_output, keywords_output, recent_works_output]
+        )
+
+        generate_button.click(
+            fn=generate_and_store, 
+            inputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output], 
+            outputs=[hypothesis_output, experiment_plan_output, hypothesis_state, experiment_plan_state]
+        )
+
+        gr.Examples(
+            examples=example_text,
+            inputs=[paper_text_input],
+            outputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output, hypothesis_output, experiment_plan_output],
+            fn=load_example_and_set_index,
+            run_on_click = True,
+            # label="⬇️ Click an example to load"
+        )
+        
+
+
+
+    ########## Phase 2 & 3:  Experiment implementation and execution ##############
+    with gr.Tab("🧪 Stage 2 & Stage 3: Experiment implementation and execution"):
+        gr.Markdown("### Interact with the ExperimentAgent")
+
+        with gr.Row():
+            with gr.Column():
+                with gr.Group():
+                    gr.Markdown("### 🌟 Generated Research Idea")
+                    with gr.Row():
+                        idea_input = gr.Textbox(label="Generated Research Hypothesis", lines=30, interactive=False)
+                        plan_input = gr.Textbox(label="Generated Experiment Plan", lines=30, interactive=False)
+        
+            with gr.Column():
+                start_exp_agnet = gr.Button("⚙️ Start / Reset ExperimentAgent", elem_classes=["agent-btn"])
+                with gr.Group():
+                    gr.Markdown("### Implementation + Execution Log")    
+                    log = gr.Textbox(label="📖 Execution Log", lines=20, interactive=False)
+                    code_display = gr.Code(label="🧑‍💻 Implementation", language="python", interactive=False)
+
+            with gr.Column():
+                response = gr.Textbox(label="🤖 ExperimentAgent Response", lines=30, interactive=False)
+                feedback = gr.Textbox(placeholder="N/A", label="🧑‍🔬 User Feedback", lines=3, interactive=True)
+                submit_button = gr.Button("Submit", elem_classes=["Submit-btn"])
+        
+        hypothesis_state.change(
+            fn=load_phase_2_inputs, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[idea_input, plan_input, code_display]
+        )
+        
+        # Start research agent
+        start_exp_agnet.click(
+            fn=start_experiment_agent, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[code_display, log, response, feedback]
+        )
+        
+        submit_button.click(
+            fn=submit_feedback,
+            inputs=[feedback, log, response],
+            outputs=[log, response, code_display, feedback]
+        )
+
+# Test
+if __name__ == "__main__":
+    step_index = 0
+    app.launch(share=True)

.history/app_20250404171127.py

@@ -0,0 +1,329 @@
+import gradio as gr
+from pathlib import Path
+from reactagent.environment import Environment
+from reactagent.agents.agent_research import ResearchAgent
+from reactagent.runner import create_parser
+from reactagent import llm
+from reactagent.users.user import User
+import os
+import json
+
+
+# Global variables to store session state
+env = None
+agent = None
+state_example = False
+state_extract = False
+state_generate = False
+state_agent = False
+state_complete = False
+index_ex = "1"
+
+example_text = [
+    "Research Paper 1: Dataset and Baseline for Automatic Student Feedback Analysis",
+    "Research Paper 2: An Empirical Study on the Impact of Code Review on Software Quality"
+]
+
+# Load example JSON file
+def load_example_data():
+    with open("example/example_data.json", "r") as json_file:
+        example_data = json.load(json_file)
+    
+    for idx in example_data.keys():
+        try:
+            file = example_data[idx]["code_init"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_init"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")
+        try:
+            file = example_data[idx]["code_final"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_final"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")     
+    return example_data
+
+example_data = load_example_data()
+
+# Function to handle the selection of an example and populate the respective fields
+def load_example(example_id):
+    global index_ex 
+    index_ex = str(example_id)
+    example = example_data[index_ex]
+    paper_text = 'Title:\t' +  example['title'] + '\n\nAbstract:\t' + example['abstract']
+    return paper_text
+
+example_text = [load_example(1), load_example(2)]
+
+# Function to handle example clicks
+def load_example_and_set_index(paper_text_input):
+    global index_ex, state_example
+    state_example = True
+    index_ex = str(example_text.index(paper_text_input) + 1)
+    paper_text =  load_example(index_ex)
+
+    return paper_text,  "", "", "", "", "", ""
+
+
+
+########## Phase 1 ##############
+
+def extract_research_elements(paper_text):
+    global state_extract, index_ex, state_example
+    if not state_example or paper_text == "":
+        return "", "", "", ""
+    state_extract = True
+    if not paper_text.strip().startswith("Title:\t" + example_data[index_ex]["title"]):
+        print("Mismatch detected.")
+        print(paper_text)
+        return "", "", "", ""
+    example = example_data[index_ex]
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return tasks, gaps, keywords, recent_works
+
+
+# Step 2: Generate Research Hypothesis and Experiment Plan
+def generate_and_store(paper_text, tasks, gaps, keywords, recent_works):
+    if (not state_extract or not state_example or paper_text == ""):
+        return "", "", "", ""
+    global state_generate, index_ex
+    state_generate = True
+    hypothesis = example_data[index_ex]['hypothesis']
+    experiment_plan = example_data[index_ex]['experiment_plan']
+    return hypothesis, experiment_plan, hypothesis, experiment_plan
+
+##########  Phase 2 & 3  ##############
+def start_experiment_agent(hypothesis, plan):
+    if (not state_extract or not state_generate or not state_example):
+        return "", "", ""
+    global state_agent, step_index, state_complete
+    state_agent = True
+    step_index = 0
+    state_complete = False
+    # predefined_message = f"Implement the following hypothesis and experiment plan:\n\nHypothesis:\n{hypothesis}\n\nExperiment Plan:\n{plan}"
+    return example_data[index_ex]['code_init'], predefined_action_log, "", ""
+
+def submit_feedback(user_feedback, history, previous_response):
+    if (not state_extract or not  state_generate or not state_agent or not state_example):
+        return "", "", ""
+    global step_index, state_complete
+    step_index += 1
+    msg = history
+    if step_index < len(process_steps):
+        msg += previous_response + "\nUser feedback:" + user_feedback + "\n\n"
+        response_info = process_steps[step_index]
+        response = info_to_message(response_info)  # Convert dictionary to formatted string
+        response += "Please provide feedback based on the history, response entries, and observation, and questions: "
+        step_index += 1
+        msg += response
+    else:
+        state_complete = True
+        response = "Agent Finished."
+        
+    return msg, response, example_data[index_ex]['code_init'] if state_complete else example_data[index_ex]['code_final'], ""
+
+def load_phase_2_inputs(hypothesis, plan):
+    return hypothesis, plan, "# Code implementation will be displayed here after Start ExperimentAgent."
+
+
+
+predefined_action_log = """
+[Reasoning]: To understand the initial structure and functionality of train.py for effective improvements.
+[Action]: Inspect Script (train.py)
+Input: {"script_name": "train.py", "start_line_number": "1", "end_line_number": "74"}
+Objective: Understand the training script, including data processing, [...]
+[Observation]: The train.py script imports [...]. Sets random seeds [...]. Defines [...] Placeholder functions [...] exist without implementation. [...]
+[Feedback]: The script structure is clear, but key functions (train_model, predict) need proper implementation for proposed model training and prediction.\n
+"""
+
+
+predefined_observation = """
+Epoch [1/10],
+Train MSE: 0.543,
+Test MSE: 0.688
+Epoch [2/10],
+Train MSE: 0.242,
+Test MSE: 0.493\n
+"""
+
+# Initialize the global step_index and history
+process_steps = [
+    {
+        "Action": "Inspect Script Lines (train.py)",
+        "Observation": (
+            "The train.py script imports necessary libraries (e.g., pandas, sklearn, torch). "
+            "Sets random seeds for reproducibility. Defines compute_metrics_for_regression function "
+            "to calculate RMSE for different dimensions. Placeholder functions train_model and "
+            "predict exist without implementations."
+        ),
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The script executed successfully. Generated embeddings using the BERT model. Completed "
+            "the training process without errors. Metrics calculation placeholders indicated areas needing implementation."
+        ),
+    },
+    {
+        "Action": "Edit Script (train.py)",
+        "Observation": (
+            "Edited train.py to separate data loading, model definition, training loop, and evaluation into distinct functions. "
+            "The edited train.py now has clearly defined functions"
+            "for data loading (load_data), model definition (build_model), "
+            "training (train_model), and evaluation (evaluate_model). Similarly, eval.py is reorganized to load the model and perform predictions efficiently."
+        ),
+    },
+    {
+        "Action": "Retrieve Model",
+        "Observation": "CNN and BiLSTM retrieved.",
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The model trained over the specified number of epochs. Training and validation loss values are recorded for each epoch, "
+            "the decrease in loss indicates improved model performance."
+        )
+    },
+    {
+        "Action": "Evaluation",
+        "Observation": predefined_observation,
+    }
+]
+def info_to_message(info):
+    msg = ""
+    for k, v in info.items():
+        if isinstance(v, dict):
+            tempv = v
+            v = ""
+            for k2, v2 in tempv.items():
+                v += f"{k2}:\n  {v2}\n"
+        v = User.indent_text(v, 2)
+        msg += '-' * 64
+        msg += '\n'
+        msg += f"{k}:\n{v}\n"
+    return msg
+
+
+def handle_example_click(example_index):
+    global index_ex
+    index_ex = example_index
+    return load_example(index_ex)  # Simply return the text to display it in the textbox
+
+# Gradio Interface
+with gr.Blocks(css=".gr-examples-label {display: none;}", theme=gr.themes.Default()) as app:
+    gr.Markdown("# MLR- Copilot: Machine Learning Research based on LLM Agents")
+    gr.Markdown("###   ")
+    gr.Markdown("## <span style='color:Orange;'> This UI is for predefined example demo only.</span>")
+    gr.Markdown("## <span style='color:Orange;'> To reproduce the results please use [Github](https://github.com/du-nlp-lab/MLR-Copilot/)</span>")
+
+
+
+    gr.Markdown("MLR-Copilot is a framework where LLMs mimic researchers’ thought processes, designed to enhance the productivity of machine learning research by automating the generation and implementation of research ideas. It begins with a research paper, autonomously generating and validating these ideas, while incorporating human feedback to help reach executable research outcomes.")
+
+
+    gr.Markdown("##    ⬇️  Click an example at bottom to start·！")
+
+
+    # Use state variables to store generated hypothesis and experiment plan
+    hypothesis_state = gr.State("")
+    experiment_plan_state = gr.State("")
+
+    ########## Phase 1:  Research Idea Generation Tab ##############
+    with gr.Tab("💡Stage 1: Research Idea Generation"):
+        gr.Markdown("### Extract Research Elements and Generate Research Ideas")
+        with gr.Row():
+            with gr.Column():
+                paper_text_input = gr.Textbox(value="", lines=10, label="📑 Research Paper Text", interactive=False)
+
+                extract_button = gr.Button("🔍 Extract Research Elements")
+                with gr.Row():
+                    tasks_output = gr.Textbox(placeholder="Research task definition", label="Research Tasks", lines=2, interactive=False)
+                    gaps_output = gr.Textbox(placeholder="Research gaps of current works", label="Research Gaps", lines=2, interactive=False)
+                    keywords_output = gr.Textbox(placeholder="Paper keywords", label="Keywords", lines=2, interactive=False)
+                    recent_works_output = gr.Textbox(placeholder="Recent works extracted from Semantic Scholar", label="Recent Works", lines=2, interactive=False)
+            with gr.Column():
+                with gr.Row():  # Move the button to the top 
+                    generate_button = gr.Button("✍️ Generate Research Hypothesis & Experiment Plan")
+                with gr.Group():
+                    gr.Markdown("### 🌟 Research Idea")
+                    with gr.Row():
+                        hypothesis_output = gr.Textbox(label="Generated Hypothesis", lines=20, interactive=False)
+                        experiment_plan_output = gr.Textbox(label="Generated Experiment Plan", lines=20, interactive=False)
+        
+    
+        # Step 1: Extract Research Elements
+        extract_button.click(
+            fn=extract_research_elements, 
+            inputs=paper_text_input, 
+            outputs=[tasks_output, gaps_output, keywords_output, recent_works_output]
+        )
+
+        generate_button.click(
+            fn=generate_and_store, 
+            inputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output], 
+            outputs=[hypothesis_output, experiment_plan_output, hypothesis_state, experiment_plan_state]
+        )
+
+        gr.Examples(
+            examples=example_text,
+            inputs=[paper_text_input],
+            outputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output, hypothesis_output, experiment_plan_output],
+            fn=load_example_and_set_index,
+            run_on_click = True,
+            # label="⬇️ Click an example to load"
+        )
+        
+
+
+
+    ########## Phase 2 & 3:  Experiment implementation and execution ##############
+    with gr.Tab("🧪 Stage 2 & Stage 3: Experiment implementation and execution"):
+        gr.Markdown("### Interact with the ExperimentAgent")
+
+        with gr.Row():
+            with gr.Column():
+                with gr.Group():
+                    gr.Markdown("### 🌟 Generated Research Idea")
+                    with gr.Row():
+                        idea_input = gr.Textbox(label="Generated Research Hypothesis", lines=30, interactive=False)
+                        plan_input = gr.Textbox(label="Generated Experiment Plan", lines=30, interactive=False)
+        
+            with gr.Column():
+                start_exp_agnet = gr.Button("⚙️ Start / Reset ExperimentAgent", elem_classes=["agent-btn"])
+                with gr.Group():
+                    gr.Markdown("### Implementation + Execution Log")    
+                    log = gr.Textbox(label="📖 Execution Log", lines=20, interactive=False)
+                    code_display = gr.Code(label="🧑‍💻 Implementation", language="python", interactive=False)
+
+            with gr.Column():
+                response = gr.Textbox(label="🤖 ExperimentAgent Response", lines=30, interactive=False)
+                feedback = gr.Textbox(placeholder="N/A", label="🧑‍🔬 User Feedback", lines=3, interactive=True)
+                submit_button = gr.Button("Submit", elem_classes=["Submit-btn"])
+        
+        hypothesis_state.change(
+            fn=load_phase_2_inputs, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[idea_input, plan_input, code_display]
+        )
+        
+        # Start research agent
+        start_exp_agnet.click(
+            fn=start_experiment_agent, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[code_display, log, response, feedback]
+        )
+        
+        submit_button.click(
+            fn=submit_feedback,
+            inputs=[feedback, log, response],
+            outputs=[log, response, code_display, feedback]
+        )
+
+# Test
+if __name__ == "__main__":
+    step_index = 0
+    app.launch(share=True)

.history/app_20250404171152.py

@@ -0,0 +1,329 @@
+import gradio as gr
+from pathlib import Path
+from reactagent.environment import Environment
+from reactagent.agents.agent_research import ResearchAgent
+from reactagent.runner import create_parser
+from reactagent import llm
+from reactagent.users.user import User
+import os
+import json
+
+
+# Global variables to store session state
+env = None
+agent = None
+state_example = False
+state_extract = False
+state_generate = False
+state_agent = False
+state_complete = False
+index_ex = "1"
+
+example_text = [
+    "Research Paper 1: Dataset and Baseline for Automatic Student Feedback Analysis",
+    "Research Paper 2: An Empirical Study on the Impact of Code Review on Software Quality"
+]
+
+# Load example JSON file
+def load_example_data():
+    with open("example/example_data.json", "r") as json_file:
+        example_data = json.load(json_file)
+    
+    for idx in example_data.keys():
+        try:
+            file = example_data[idx]["code_init"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_init"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")
+        try:
+            file = example_data[idx]["code_final"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_final"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")     
+    return example_data
+
+example_data = load_example_data()
+
+# Function to handle the selection of an example and populate the respective fields
+def load_example(example_id):
+    global index_ex 
+    index_ex = str(example_id)
+    example = example_data[index_ex]
+    paper_text = 'Title:\t' +  example['title'] + '\n\nAbstract:\t' + example['abstract']
+    return paper_text
+
+example_text = [load_example(1), load_example(2)]
+
+# Function to handle example clicks
+def load_example_and_set_index(paper_text_input):
+    global index_ex, state_example
+    state_example = True
+    index_ex = str(example_text.index(paper_text_input) + 1)
+    paper_text =  load_example(index_ex)
+
+    return paper_text,  "", "", "", "", "", ""
+
+
+
+########## Phase 1 ##############
+
+def extract_research_elements(paper_text):
+    global state_extract, index_ex, state_example
+    if not state_example or paper_text == "":
+        return "", "", "", ""
+    state_extract = True
+    if not paper_text.strip().startswith("Title:\t" + example_data[index_ex]["title"]):
+        print("Mismatch detected.")
+        print(paper_text)
+        return "", "", "", ""
+    example = example_data[index_ex]
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return tasks, gaps, keywords, recent_works
+
+
+# Step 2: Generate Research Hypothesis and Experiment Plan
+def generate_and_store(paper_text, tasks, gaps, keywords, recent_works):
+    if (not state_extract or not state_example or paper_text == ""):
+        return "", "", "", ""
+    global state_generate, index_ex
+    state_generate = True
+    hypothesis = example_data[index_ex]['hypothesis']
+    experiment_plan = example_data[index_ex]['experiment_plan']
+    return hypothesis, experiment_plan, hypothesis, experiment_plan
+
+##########  Phase 2 & 3  ##############
+def start_experiment_agent(hypothesis, plan):
+    if (not state_extract or not state_generate or not state_example):
+        return "", "", ""
+    global state_agent, step_index, state_complete
+    state_agent = True
+    step_index = 0
+    state_complete = False
+    # predefined_message = f"Implement the following hypothesis and experiment plan:\n\nHypothesis:\n{hypothesis}\n\nExperiment Plan:\n{plan}"
+    return example_data[index_ex]['code_init'], predefined_action_log, "", ""
+
+def submit_feedback(user_feedback, history, previous_response):
+    if (not state_extract or not  state_generate or not state_agent or not state_example):
+        return "", "", ""
+    global step_index, state_complete
+    step_index += 1
+    msg = history
+    if step_index < len(process_steps):
+        msg += previous_response + "\nUser feedback:" + user_feedback + "\n\n"
+        response_info = process_steps[step_index]
+        response = info_to_message(response_info)  # Convert dictionary to formatted string
+        response += "Please provide feedback based on the history, response entries, and observation, and questions: "
+        step_index += 1
+        msg += response
+    else:
+        state_complete = True
+        response = "Agent Finished."
+        
+    return msg, response, example_data[index_ex]['code_init'] if state_complete else example_data[index_ex]['code_final'], ""
+
+def load_phase_2_inputs(hypothesis, plan):
+    return hypothesis, plan, "# Code implementation will be displayed here after Start ExperimentAgent."
+
+
+
+predefined_action_log = """
+[Reasoning]: To understand the initial structure and functionality of train.py for effective improvements.
+[Action]: Inspect Script (train.py)
+Input: {"script_name": "train.py", "start_line_number": "1", "end_line_number": "74"}
+Objective: Understand the training script, including data processing, [...]
+[Observation]: The train.py script imports [...]. Sets random seeds [...]. Defines [...] Placeholder functions [...] exist without implementation. [...]
+[Feedback]: The script structure is clear, but key functions (train_model, predict) need proper implementation for proposed model training and prediction.\n
+"""
+
+
+predefined_observation = """
+Epoch [1/10],
+Train MSE: 0.543,
+Test MSE: 0.688
+Epoch [2/10],
+Train MSE: 0.242,
+Test MSE: 0.493\n
+"""
+
+# Initialize the global step_index and history
+process_steps = [
+    {
+        "Action": "Inspect Script Lines (train.py)",
+        "Observation": (
+            "The train.py script imports necessary libraries (e.g., pandas, sklearn, torch). "
+            "Sets random seeds for reproducibility. Defines compute_metrics_for_regression function "
+            "to calculate RMSE for different dimensions. Placeholder functions train_model and "
+            "predict exist without implementations."
+        ),
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The script executed successfully. Generated embeddings using the BERT model. Completed "
+            "the training process without errors. Metrics calculation placeholders indicated areas needing implementation."
+        ),
+    },
+    {
+        "Action": "Edit Script (train.py)",
+        "Observation": (
+            "Edited train.py to separate data loading, model definition, training loop, and evaluation into distinct functions. "
+            "The edited train.py now has clearly defined functions"
+            "for data loading (load_data), model definition (build_model), "
+            "training (train_model), and evaluation (evaluate_model). Similarly, eval.py is reorganized to load the model and perform predictions efficiently."
+        ),
+    },
+    {
+        "Action": "Retrieve Model",
+        "Observation": "CNN and BiLSTM retrieved.",
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The model trained over the specified number of epochs. Training and validation loss values are recorded for each epoch, "
+            "the decrease in loss indicates improved model performance."
+        )
+    },
+    {
+        "Action": "Evaluation",
+        "Observation": predefined_observation,
+    }
+]
+def info_to_message(info):
+    msg = ""
+    for k, v in info.items():
+        if isinstance(v, dict):
+            tempv = v
+            v = ""
+            for k2, v2 in tempv.items():
+                v += f"{k2}:\n  {v2}\n"
+        v = User.indent_text(v, 2)
+        msg += '-' * 64
+        msg += '\n'
+        msg += f"{k}:\n{v}\n"
+    return msg
+
+
+def handle_example_click(example_index):
+    global index_ex
+    index_ex = example_index
+    return load_example(index_ex)  # Simply return the text to display it in the textbox
+
+# Gradio Interface
+with gr.Blocks(css=".gr-examples-label {display: none;}", theme=gr.themes.Default()) as app:
+    gr.Markdown("# MLR- Copilot: Machine Learning Research based on LLM Agents")
+    gr.Markdown("###   ")
+    gr.Markdown("## <span style='color:Orange;'> This UI is for predefined example demo only.</span>")
+    gr.Markdown("## <span style='color:Orange;'> To reproduce the results please use [Github](https://github.com/du-nlp-lab/MLR-Copilot/)</span>")
+
+
+
+    gr.Markdown("MLR-Copilot is a framework where LLMs mimic researchers’ thought processes, designed to enhance the productivity of machine learning research by automating the generation and implementation of research ideas. It begins with a research paper, autonomously generating and validating these ideas, while incorporating human feedback to help reach executable research outcomes.")
+
+
+    gr.Markdown("##    ⬇️  Click an example at bottom to start！")
+
+
+    # Use state variables to store generated hypothesis and experiment plan
+    hypothesis_state = gr.State("")
+    experiment_plan_state = gr.State("")
+
+    ########## Phase 1:  Research Idea Generation Tab ##############
+    with gr.Tab("💡Stage 1: Research Idea Generation"):
+        gr.Markdown("### Extract Research Elements and Generate Research Ideas")
+        with gr.Row():
+            with gr.Column():
+                paper_text_input = gr.Textbox(value="", lines=10, label="📑 Research Paper Text", interactive=False)
+
+                extract_button = gr.Button("🔍 Extract Research Elements")
+                with gr.Row():
+                    tasks_output = gr.Textbox(placeholder="Research task definition", label="Research Tasks", lines=2, interactive=False)
+                    gaps_output = gr.Textbox(placeholder="Research gaps of current works", label="Research Gaps", lines=2, interactive=False)
+                    keywords_output = gr.Textbox(placeholder="Paper keywords", label="Keywords", lines=2, interactive=False)
+                    recent_works_output = gr.Textbox(placeholder="Recent works extracted from Semantic Scholar", label="Recent Works", lines=2, interactive=False)
+            with gr.Column():
+                with gr.Row():  # Move the button to the top 
+                    generate_button = gr.Button("✍️ Generate Research Hypothesis & Experiment Plan")
+                with gr.Group():
+                    gr.Markdown("### 🌟 Research Idea")
+                    with gr.Row():
+                        hypothesis_output = gr.Textbox(label="Generated Hypothesis", lines=20, interactive=False)
+                        experiment_plan_output = gr.Textbox(label="Generated Experiment Plan", lines=20, interactive=False)
+        
+    
+        # Step 1: Extract Research Elements
+        extract_button.click(
+            fn=extract_research_elements, 
+            inputs=paper_text_input, 
+            outputs=[tasks_output, gaps_output, keywords_output, recent_works_output]
+        )
+
+        generate_button.click(
+            fn=generate_and_store, 
+            inputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output], 
+            outputs=[hypothesis_output, experiment_plan_output, hypothesis_state, experiment_plan_state]
+        )
+
+        gr.Examples(
+            examples=example_text,
+            inputs=[paper_text_input],
+            outputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output, hypothesis_output, experiment_plan_output],
+            fn=load_example_and_set_index,
+            run_on_click = True,
+            # label="⬇️ Click an example to load"
+        )
+        
+
+
+
+    ########## Phase 2 & 3:  Experiment implementation and execution ##############
+    with gr.Tab("🧪 Stage 2 & Stage 3: Experiment implementation and execution"):
+        gr.Markdown("### Interact with the ExperimentAgent")
+
+        with gr.Row():
+            with gr.Column():
+                with gr.Group():
+                    gr.Markdown("### 🌟 Generated Research Idea")
+                    with gr.Row():
+                        idea_input = gr.Textbox(label="Generated Research Hypothesis", lines=30, interactive=False)
+                        plan_input = gr.Textbox(label="Generated Experiment Plan", lines=30, interactive=False)
+        
+            with gr.Column():
+                start_exp_agnet = gr.Button("⚙️ Start / Reset ExperimentAgent", elem_classes=["agent-btn"])
+                with gr.Group():
+                    gr.Markdown("### Implementation + Execution Log")    
+                    log = gr.Textbox(label="📖 Execution Log", lines=20, interactive=False)
+                    code_display = gr.Code(label="🧑‍💻 Implementation", language="python", interactive=False)
+
+            with gr.Column():
+                response = gr.Textbox(label="🤖 ExperimentAgent Response", lines=30, interactive=False)
+                feedback = gr.Textbox(placeholder="N/A", label="🧑‍🔬 User Feedback", lines=3, interactive=True)
+                submit_button = gr.Button("Submit", elem_classes=["Submit-btn"])
+        
+        hypothesis_state.change(
+            fn=load_phase_2_inputs, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[idea_input, plan_input, code_display]
+        )
+        
+        # Start research agent
+        start_exp_agnet.click(
+            fn=start_experiment_agent, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[code_display, log, response, feedback]
+        )
+        
+        submit_button.click(
+            fn=submit_feedback,
+            inputs=[feedback, log, response],
+            outputs=[log, response, code_display, feedback]
+        )
+
+# Test
+if __name__ == "__main__":
+    step_index = 0
+    app.launch(share=True)

.history/app_20250404171213.py

@@ -0,0 +1,329 @@
+import gradio as gr
+from pathlib import Path
+from reactagent.environment import Environment
+from reactagent.agents.agent_research import ResearchAgent
+from reactagent.runner import create_parser
+from reactagent import llm
+from reactagent.users.user import User
+import os
+import json
+
+
+# Global variables to store session state
+env = None
+agent = None
+state_example = False
+state_extract = False
+state_generate = False
+state_agent = False
+state_complete = False
+index_ex = "1"
+
+example_text = [
+    "Research Paper 1: Dataset and Baseline for Automatic Student Feedback Analysis",
+    "Research Paper 2: An Empirical Study on the Impact of Code Review on Software Quality"
+]
+
+# Load example JSON file
+def load_example_data():
+    with open("example/example_data.json", "r") as json_file:
+        example_data = json.load(json_file)
+    
+    for idx in example_data.keys():
+        try:
+            file = example_data[idx]["code_init"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_init"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")
+        try:
+            file = example_data[idx]["code_final"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_final"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")     
+    return example_data
+
+example_data = load_example_data()
+
+# Function to handle the selection of an example and populate the respective fields
+def load_example(example_id):
+    global index_ex 
+    index_ex = str(example_id)
+    example = example_data[index_ex]
+    paper_text = 'Title:\t' +  example['title'] + '\n\nAbstract:\t' + example['abstract']
+    return paper_text
+
+example_text = [load_example(1), load_example(2)]
+
+# Function to handle example clicks
+def load_example_and_set_index(paper_text_input):
+    global index_ex, state_example
+    state_example = True
+    index_ex = str(example_text.index(paper_text_input) + 1)
+    paper_text =  load_example(index_ex)
+
+    return paper_text,  "", "", "", "", "", ""
+
+
+
+########## Phase 1 ##############
+
+def extract_research_elements(paper_text):
+    global state_extract, index_ex, state_example
+    if not state_example or paper_text == "":
+        return "", "", "", ""
+    state_extract = True
+    if not paper_text.strip().startswith("Title:\t" + example_data[index_ex]["title"]):
+        print("Mismatch detected.")
+        print(paper_text)
+        return "", "", "", ""
+    example = example_data[index_ex]
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return tasks, gaps, keywords, recent_works
+
+
+# Step 2: Generate Research Hypothesis and Experiment Plan
+def generate_and_store(paper_text, tasks, gaps, keywords, recent_works):
+    if (not state_extract or not state_example or paper_text == ""):
+        return "", "", "", ""
+    global state_generate, index_ex
+    state_generate = True
+    hypothesis = example_data[index_ex]['hypothesis']
+    experiment_plan = example_data[index_ex]['experiment_plan']
+    return hypothesis, experiment_plan, hypothesis, experiment_plan
+
+##########  Phase 2 & 3  ##############
+def start_experiment_agent(hypothesis, plan):
+    if (not state_extract or not state_generate or not state_example):
+        return "", "", ""
+    global state_agent, step_index, state_complete
+    state_agent = True
+    step_index = 0
+    state_complete = False
+    # predefined_message = f"Implement the following hypothesis and experiment plan:\n\nHypothesis:\n{hypothesis}\n\nExperiment Plan:\n{plan}"
+    return example_data[index_ex]['code_init'], predefined_action_log, "", ""
+
+def submit_feedback(user_feedback, history, previous_response):
+    if (not state_extract or not  state_generate or not state_agent or not state_example):
+        return "", "", ""
+    global step_index, state_complete
+    step_index += 1
+    msg = history
+    if step_index < len(process_steps):
+        msg += previous_response + "\nUser feedback:" + user_feedback + "\n\n"
+        response_info = process_steps[step_index]
+        response = info_to_message(response_info)  # Convert dictionary to formatted string
+        response += "Please provide feedback based on the history, response entries, and observation, and questions: "
+        step_index += 1
+        msg += response
+    else:
+        state_complete = True
+        response = "Agent Finished."
+        
+    return msg, response, example_data[index_ex]['code_init'] if state_complete else example_data[index_ex]['code_final'], ""
+
+def load_phase_2_inputs(hypothesis, plan):
+    return hypothesis, plan, "# Code implementation will be displayed here after Start ExperimentAgent."
+
+
+
+predefined_action_log = """
+[Reasoning]: To understand the initial structure and functionality of train.py for effective improvements.
+[Action]: Inspect Script (train.py)
+Input: {"script_name": "train.py", "start_line_number": "1", "end_line_number": "74"}
+Objective: Understand the training script, including data processing, [...]
+[Observation]: The train.py script imports [...]. Sets random seeds [...]. Defines [...] Placeholder functions [...] exist without implementation. [...]
+[Feedback]: The script structure is clear, but key functions (train_model, predict) need proper implementation for proposed model training and prediction.\n
+"""
+
+
+predefined_observation = """
+Epoch [1/10],
+Train MSE: 0.543,
+Test MSE: 0.688
+Epoch [2/10],
+Train MSE: 0.242,
+Test MSE: 0.493\n
+"""
+
+# Initialize the global step_index and history
+process_steps = [
+    {
+        "Action": "Inspect Script Lines (train.py)",
+        "Observation": (
+            "The train.py script imports necessary libraries (e.g., pandas, sklearn, torch). "
+            "Sets random seeds for reproducibility. Defines compute_metrics_for_regression function "
+            "to calculate RMSE for different dimensions. Placeholder functions train_model and "
+            "predict exist without implementations."
+        ),
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The script executed successfully. Generated embeddings using the BERT model. Completed "
+            "the training process without errors. Metrics calculation placeholders indicated areas needing implementation."
+        ),
+    },
+    {
+        "Action": "Edit Script (train.py)",
+        "Observation": (
+            "Edited train.py to separate data loading, model definition, training loop, and evaluation into distinct functions. "
+            "The edited train.py now has clearly defined functions"
+            "for data loading (load_data), model definition (build_model), "
+            "training (train_model), and evaluation (evaluate_model). Similarly, eval.py is reorganized to load the model and perform predictions efficiently."
+        ),
+    },
+    {
+        "Action": "Retrieve Model",
+        "Observation": "CNN and BiLSTM retrieved.",
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The model trained over the specified number of epochs. Training and validation loss values are recorded for each epoch, "
+            "the decrease in loss indicates improved model performance."
+        )
+    },
+    {
+        "Action": "Evaluation",
+        "Observation": predefined_observation,
+    }
+]
+def info_to_message(info):
+    msg = ""
+    for k, v in info.items():
+        if isinstance(v, dict):
+            tempv = v
+            v = ""
+            for k2, v2 in tempv.items():
+                v += f"{k2}:\n  {v2}\n"
+        v = User.indent_text(v, 2)
+        msg += '-' * 64
+        msg += '\n'
+        msg += f"{k}:\n{v}\n"
+    return msg
+
+
+def handle_example_click(example_index):
+    global index_ex
+    index_ex = example_index
+    return load_example(index_ex)  # Simply return the text to display it in the textbox
+
+# Gradio Interface
+with gr.Blocks(css=".gr-examples-label {display: none;}", theme=gr.themes.Default()) as app:
+    gr.Markdown("# MLR- Copilot: Machine Learning Research based on LLM Agents")
+    gr.Markdown("###   ")
+    gr.Markdown("## <span style='color:Orange;'> This UI is for predefined example demo only.</span>")
+    gr.Markdown("## <span style='color:Orange;'> To reproduce the results please use [Github](https://github.com/du-nlp-lab/MLR-Copilot/)</span>")
+
+
+
+    gr.Markdown("MLR-Copilot is a framework where LLMs mimic researchers’ thought processes, designed to enhance the productivity of machine learning research by automating the generation and implementation of research ideas. It begins with a research paper, autonomously generating and validating these ideas, while incorporating human feedback to help reach executable research outcomes.")
+
+
+    gr.Markdown("##    ⬇️  Click an example at bottom to start！")
+
+
+    # Use state variables to store generated hypothesis and experiment plan
+    hypothesis_state = gr.State("")
+    experiment_plan_state = gr.State("")
+
+    ########## Phase 1:  Research Idea Generation Tab ##############
+    with gr.Tab("💡Stage 1: Research Idea Generation"):
+        gr.Markdown("### Extract Research Elements and Generate Research Ideas")
+        with gr.Row():
+            with gr.Column():
+                paper_text_input = gr.Textbox(value="", lines=10, label="📑 Research Paper Text", interactive=False)
+
+                extract_button = gr.Button("🔍 Extract Research Elements")
+                with gr.Row():
+                    tasks_output = gr.Textbox(placeholder="Research task definition", label="Research Tasks", lines=2, interactive=False)
+                    gaps_output = gr.Textbox(placeholder="Research gaps of current works", label="Research Gaps", lines=2, interactive=False)
+                    keywords_output = gr.Textbox(placeholder="Paper keywords", label="Keywords", lines=2, interactive=False)
+                    recent_works_output = gr.Textbox(placeholder="Recent works extracted from Semantic Scholar", label="Recent Works", lines=2, interactive=False)
+            with gr.Column():
+                with gr.Row():  # Move the button to the top 
+                    generate_button = gr.Button("✍️ Generate Research Hypothesis & Experiment Plan")
+                with gr.Group():
+                    gr.Markdown("### 🌟 Research Idea")
+                    with gr.Row():
+                        hypothesis_output = gr.Textbox(label="Generated Hypothesis", lines=20, interactive=False)
+                        experiment_plan_output = gr.Textbox(label="Generated Experiment Plan", lines=20, interactive=False)
+        
+    
+        # Step 1: Extract Research Elements
+        extract_button.click(
+            fn=extract_research_elements, 
+            inputs=paper_text_input, 
+            outputs=[tasks_output, gaps_output, keywords_output, recent_works_output]
+        )
+
+        generate_button.click(
+            fn=generate_and_store, 
+            inputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output], 
+            outputs=[hypothesis_output, experiment_plan_output, hypothesis_state, experiment_plan_state]
+        )
+
+        gr.Examples(
+            examples=example_text,
+            inputs=[paper_text_input],
+            outputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output, hypothesis_output, experiment_plan_output],
+            fn=load_example_and_set_index,
+            run_on_click = True,
+            # label="⬇️ Click an example to load"
+        )
+        
+
+
+
+    ########## Phase 2 & 3:  Experiment implementation and execution ##############
+    with gr.Tab("🧪 Stage 2 & Stage 3: Experiment implementation and execution"):
+        gr.Markdown("### Interact with the ExperimentAgent")
+
+        with gr.Row():
+            with gr.Column():
+                with gr.Group():
+                    gr.Markdown("### 🌟 Generated Research Idea")
+                    with gr.Row():
+                        idea_input = gr.Textbox(label="Generated Research Hypothesis", lines=30, interactive=False)
+                        plan_input = gr.Textbox(label="Generated Experiment Plan", lines=30, interactive=False)
+        
+            with gr.Column():
+                start_exp_agnet = gr.Button("⚙️ Start / Reset ExperimentAgent", elem_classes=["agent-btn"])
+                with gr.Group():
+                    gr.Markdown("### Implementation + Execution Log")    
+                    log = gr.Textbox(label="📖 Execution Log", lines=20, interactive=False)
+                    code_display = gr.Code(label="🧑‍💻 Implementation", language="python", interactive=False)
+
+            with gr.Column():
+                response = gr.Textbox(label="🤖 ExperimentAgent Response", lines=30, interactive=False)
+                feedback = gr.Textbox(placeholder="N/A", label="🧑‍🔬 User Feedback", lines=3, interactive=True)
+                submit_button = gr.Button("Submit", elem_classes=["Submit-btn"])
+        
+        hypothesis_state.change(
+            fn=load_phase_2_inputs, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[idea_input, plan_input, code_display]
+        )
+        
+        # Start research agent
+        start_exp_agnet.click(
+            fn=start_experiment_agent, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[code_display, log, response, feedback]
+        )
+        
+        submit_button.click(
+            fn=submit_feedback,
+            inputs=[feedback, log, response],
+            outputs=[log, response, code_display, feedback]
+        )
+
+# Test
+if __name__ == "__main__":
+    step_index = 0
+    app.launch(share=True)

.history/app_20250404171307.py

@@ -0,0 +1,329 @@
+import gradio as gr
+from pathlib import Path
+from reactagent.environment import Environment
+from reactagent.agents.agent_research import ResearchAgent
+from reactagent.runner import create_parser
+from reactagent import llm
+from reactagent.users.user import User
+import os
+import json
+
+
+# Global variables to store session state
+env = None
+agent = None
+state_example = False
+state_extract = False
+state_generate = False
+state_agent = False
+state_complete = False
+index_ex = "1"
+
+example_text = [
+    "Research Paper 1: Dataset and Baseline for Automatic Student Feedback Analysis",
+    "Research Paper 2: An Empirical Study on the Impact of Code Review on Software Quality"
+]
+
+# Load example JSON file
+def load_example_data():
+    with open("example/example_data.json", "r") as json_file:
+        example_data = json.load(json_file)
+    
+    for idx in example_data.keys():
+        try:
+            file = example_data[idx]["code_init"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_init"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")
+        try:
+            file = example_data[idx]["code_final"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_final"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")     
+    return example_data
+
+example_data = load_example_data()
+
+# Function to handle the selection of an example and populate the respective fields
+def load_example(example_id):
+    global index_ex 
+    index_ex = str(example_id)
+    example = example_data[index_ex]
+    paper_text = 'Title:\t' +  example['title'] + '\n\nAbstract:\t' + example['abstract']
+    return paper_text
+
+example_text = [load_example(1), load_example(2)]
+
+# Function to handle example clicks
+def load_example_and_set_index(paper_text_input):
+    global index_ex, state_example
+    state_example = True
+    index_ex = str(example_text.index(paper_text_input) + 1)
+    paper_text =  load_example(index_ex)
+
+    return paper_text,  "", "", "", "", "", ""
+
+
+
+########## Phase 1 ##############
+
+def extract_research_elements(paper_text):
+    global state_extract, index_ex, state_example
+    if not state_example or paper_text == "":
+        return "", "", "", ""
+    state_extract = True
+    if not paper_text.strip().startswith("Title:\t" + example_data[index_ex]["title"]):
+        print("Mismatch detected.")
+        print(paper_text)
+        return "", "", "", ""
+    example = example_data[index_ex]
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return tasks, gaps, keywords, recent_works
+
+
+# Step 2: Generate Research Hypothesis and Experiment Plan
+def generate_and_store(paper_text, tasks, gaps, keywords, recent_works):
+    if (not state_extract or not state_example or paper_text == ""):
+        return "", "", "", ""
+    global state_generate, index_ex
+    state_generate = True
+    hypothesis = example_data[index_ex]['hypothesis']
+    experiment_plan = example_data[index_ex]['experiment_plan']
+    return hypothesis, experiment_plan, hypothesis, experiment_plan
+
+##########  Phase 2 & 3  ##############
+def start_experiment_agent(hypothesis, plan):
+    if (not state_extract or not state_generate or not state_example):
+        return "", "", ""
+    global state_agent, step_index, state_complete
+    state_agent = True
+    step_index = 0
+    state_complete = False
+    # predefined_message = f"Implement the following hypothesis and experiment plan:\n\nHypothesis:\n{hypothesis}\n\nExperiment Plan:\n{plan}"
+    return example_data[index_ex]['code_init'], predefined_action_log, "", ""
+
+def submit_feedback(user_feedback, history, previous_response):
+    if (not state_extract or not  state_generate or not state_agent or not state_example):
+        return "", "", ""
+    global step_index, state_complete
+    step_index += 1
+    msg = history
+    if step_index < len(process_steps):
+        msg += previous_response + "\nUser feedback:" + user_feedback + "\n\n"
+        response_info = process_steps[step_index]
+        response = info_to_message(response_info)  # Convert dictionary to formatted string
+        response += "Please provide feedback based on the history, response entries, and observation, and questions: "
+        step_index += 1
+        msg += response
+    else:
+        state_complete = True
+        response = "Agent Finished."
+        
+    return msg, response, example_data[index_ex]['code_init'] if state_complete else example_data[index_ex]['code_final'], ""
+
+def load_phase_2_inputs(hypothesis, plan):
+    return hypothesis, plan, "# Code implementation will be displayed here after Start ExperimentAgent."
+
+
+
+predefined_action_log = """
+[Reasoning]: To understand the initial structure and functionality of train.py for effective improvements.
+[Action]: Inspect Script (train.py)
+Input: {"script_name": "train.py", "start_line_number": "1", "end_line_number": "74"}
+Objective: Understand the training script, including data processing, [...]
+[Observation]: The train.py script imports [...]. Sets random seeds [...]. Defines [...] Placeholder functions [...] exist without implementation. [...]
+[Feedback]: The script structure is clear, but key functions (train_model, predict) need proper implementation for proposed model training and prediction.\n
+"""
+
+
+predefined_observation = """
+Epoch [1/10],
+Train MSE: 0.543,
+Test MSE: 0.688
+Epoch [2/10],
+Train MSE: 0.242,
+Test MSE: 0.493\n
+"""
+
+# Initialize the global step_index and history
+process_steps = [
+    {
+        "Action": "Inspect Script Lines (train.py)",
+        "Observation": (
+            "The train.py script imports necessary libraries (e.g., pandas, sklearn, torch). "
+            "Sets random seeds for reproducibility. Defines compute_metrics_for_regression function "
+            "to calculate RMSE for different dimensions. Placeholder functions train_model and "
+            "predict exist without implementations."
+        ),
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The script executed successfully. Generated embeddings using the BERT model. Completed "
+            "the training process without errors. Metrics calculation placeholders indicated areas needing implementation."
+        ),
+    },
+    {
+        "Action": "Edit Script (train.py)",
+        "Observation": (
+            "Edited train.py to separate data loading, model definition, training loop, and evaluation into distinct functions. "
+            "The edited train.py now has clearly defined functions"
+            "for data loading (load_data), model definition (build_model), "
+            "training (train_model), and evaluation (evaluate_model). Similarly, eval.py is reorganized to load the model and perform predictions efficiently."
+        ),
+    },
+    {
+        "Action": "Retrieve Model",
+        "Observation": "CNN and BiLSTM retrieved.",
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The model trained over the specified number of epochs. Training and validation loss values are recorded for each epoch, "
+            "the decrease in loss indicates improved model performance."
+        )
+    },
+    {
+        "Action": "Evaluation",
+        "Observation": predefined_observation,
+    }
+]
+def info_to_message(info):
+    msg = ""
+    for k, v in info.items():
+        if isinstance(v, dict):
+            tempv = v
+            v = ""
+            for k2, v2 in tempv.items():
+                v += f"{k2}:\n  {v2}\n"
+        v = User.indent_text(v, 2)
+        msg += '-' * 64
+        msg += '\n'
+        msg += f"{k}:\n{v}\n"
+    return msg
+
+
+def handle_example_click(example_index):
+    global index_ex
+    index_ex = example_index
+    return load_example(index_ex)  # Simply return the text to display it in the textbox
+
+# Gradio Interface
+with gr.Blocks(css=".gr-examples-label {display: none;}", theme=gr.themes.Default()) as app:
+    gr.Markdown("# MLR- Copilot: Machine Learning Research based on LLM Agents")
+    gr.Markdown("###   ")
+    gr.Markdown("## <span style='color:Orange;'> This UI is for predefined example demo only.</span>")
+    gr.Markdown("## <span style='color:Orange;'> To reproduce the results please use [Github](https://github.com/du-nlp-lab/MLR-Copilot/)</span>")
+
+
+
+    gr.Markdown("MLR-Copilot is a framework where LLMs mimic researchers’ thought processes, designed to enhance the productivity of machine learning research by automating the generation and implementation of research ideas. It begins with a research paper, autonomously generating and validating these ideas, while incorporating human feedback to help reach executable research outcomes.")
+
+
+    gr.Markdown("##    ⬇️  Click an example at bottom to start！")
+
+
+    # Use state variables to store generated hypothesis and experiment plan
+    hypothesis_state = gr.State("")
+    experiment_plan_state = gr.State("")
+
+    ########## Phase 1:  Research Idea Generation Tab ##############
+    with gr.Tab("💡Stage 1: Research Idea Generation"):
+        gr.Markdown("### Extract Research Elements and Generate Research Ideas")
+        with gr.Row():
+            with gr.Column():
+                paper_text_input = gr.Textbox(value="", lines=10, label="📑 Research Paper Text", interactive=False)
+
+                # extract_button = gr.Button("🔍 Extract Research Elements")
+                # with gr.Row():
+                #     tasks_output = gr.Textbox(placeholder="Research task definition", label="Research Tasks", lines=2, interactive=False)
+                #     gaps_output = gr.Textbox(placeholder="Research gaps of current works", label="Research Gaps", lines=2, interactive=False)
+                #     keywords_output = gr.Textbox(placeholder="Paper keywords", label="Keywords", lines=2, interactive=False)
+                #     recent_works_output = gr.Textbox(placeholder="Recent works extracted from Semantic Scholar", label="Recent Works", lines=2, interactive=False)
+            with gr.Column():
+                with gr.Row():  # Move the button to the top 
+                    generate_button = gr.Button("✍️ Generate Research Hypothesis & Experiment Plan")
+                with gr.Group():
+                    gr.Markdown("### 🌟 Research Idea")
+                    with gr.Row():
+                        hypothesis_output = gr.Textbox(label="Generated Hypothesis", lines=20, interactive=False)
+                        experiment_plan_output = gr.Textbox(label="Generated Experiment Plan", lines=20, interactive=False)
+        
+    
+        # Step 1: Extract Research Elements
+        extract_button.click(
+            fn=extract_research_elements, 
+            inputs=paper_text_input, 
+            outputs=[tasks_output, gaps_output, keywords_output, recent_works_output]
+        )
+
+        generate_button.click(
+            fn=generate_and_store, 
+            inputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output], 
+            outputs=[hypothesis_output, experiment_plan_output, hypothesis_state, experiment_plan_state]
+        )
+
+        gr.Examples(
+            examples=example_text,
+            inputs=[paper_text_input],
+            outputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output, hypothesis_output, experiment_plan_output],
+            fn=load_example_and_set_index,
+            run_on_click = True,
+            # label="⬇️ Click an example to load"
+        )
+        
+
+
+
+    ########## Phase 2 & 3:  Experiment implementation and execution ##############
+    with gr.Tab("🧪 Stage 2 & Stage 3: Experiment implementation and execution"):
+        gr.Markdown("### Interact with the ExperimentAgent")
+
+        with gr.Row():
+            with gr.Column():
+                with gr.Group():
+                    gr.Markdown("### 🌟 Generated Research Idea")
+                    with gr.Row():
+                        idea_input = gr.Textbox(label="Generated Research Hypothesis", lines=30, interactive=False)
+                        plan_input = gr.Textbox(label="Generated Experiment Plan", lines=30, interactive=False)
+        
+            with gr.Column():
+                start_exp_agnet = gr.Button("⚙️ Start / Reset ExperimentAgent", elem_classes=["agent-btn"])
+                with gr.Group():
+                    gr.Markdown("### Implementation + Execution Log")    
+                    log = gr.Textbox(label="📖 Execution Log", lines=20, interactive=False)
+                    code_display = gr.Code(label="🧑‍💻 Implementation", language="python", interactive=False)
+
+            with gr.Column():
+                response = gr.Textbox(label="🤖 ExperimentAgent Response", lines=30, interactive=False)
+                feedback = gr.Textbox(placeholder="N/A", label="🧑‍🔬 User Feedback", lines=3, interactive=True)
+                submit_button = gr.Button("Submit", elem_classes=["Submit-btn"])
+        
+        hypothesis_state.change(
+            fn=load_phase_2_inputs, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[idea_input, plan_input, code_display]
+        )
+        
+        # Start research agent
+        start_exp_agnet.click(
+            fn=start_experiment_agent, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[code_display, log, response, feedback]
+        )
+        
+        submit_button.click(
+            fn=submit_feedback,
+            inputs=[feedback, log, response],
+            outputs=[log, response, code_display, feedback]
+        )
+
+# Test
+if __name__ == "__main__":
+    step_index = 0
+    app.launch(share=True)

.history/app_20250404171308.py

@@ -0,0 +1,329 @@
+import gradio as gr
+from pathlib import Path
+from reactagent.environment import Environment
+from reactagent.agents.agent_research import ResearchAgent
+from reactagent.runner import create_parser
+from reactagent import llm
+from reactagent.users.user import User
+import os
+import json
+
+
+# Global variables to store session state
+env = None
+agent = None
+state_example = False
+state_extract = False
+state_generate = False
+state_agent = False
+state_complete = False
+index_ex = "1"
+
+example_text = [
+    "Research Paper 1: Dataset and Baseline for Automatic Student Feedback Analysis",
+    "Research Paper 2: An Empirical Study on the Impact of Code Review on Software Quality"
+]
+
+# Load example JSON file
+def load_example_data():
+    with open("example/example_data.json", "r") as json_file:
+        example_data = json.load(json_file)
+    
+    for idx in example_data.keys():
+        try:
+            file = example_data[idx]["code_init"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_init"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")
+        try:
+            file = example_data[idx]["code_final"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_final"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")     
+    return example_data
+
+example_data = load_example_data()
+
+# Function to handle the selection of an example and populate the respective fields
+def load_example(example_id):
+    global index_ex 
+    index_ex = str(example_id)
+    example = example_data[index_ex]
+    paper_text = 'Title:\t' +  example['title'] + '\n\nAbstract:\t' + example['abstract']
+    return paper_text
+
+example_text = [load_example(1), load_example(2)]
+
+# Function to handle example clicks
+def load_example_and_set_index(paper_text_input):
+    global index_ex, state_example
+    state_example = True
+    index_ex = str(example_text.index(paper_text_input) + 1)
+    paper_text =  load_example(index_ex)
+
+    return paper_text,  "", "", "", "", "", ""
+
+
+
+########## Phase 1 ##############
+
+def extract_research_elements(paper_text):
+    global state_extract, index_ex, state_example
+    if not state_example or paper_text == "":
+        return "", "", "", ""
+    state_extract = True
+    if not paper_text.strip().startswith("Title:\t" + example_data[index_ex]["title"]):
+        print("Mismatch detected.")
+        print(paper_text)
+        return "", "", "", ""
+    example = example_data[index_ex]
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return tasks, gaps, keywords, recent_works
+
+
+# Step 2: Generate Research Hypothesis and Experiment Plan
+def generate_and_store(paper_text, tasks, gaps, keywords, recent_works):
+    if (not state_extract or not state_example or paper_text == ""):
+        return "", "", "", ""
+    global state_generate, index_ex
+    state_generate = True
+    hypothesis = example_data[index_ex]['hypothesis']
+    experiment_plan = example_data[index_ex]['experiment_plan']
+    return hypothesis, experiment_plan, hypothesis, experiment_plan
+
+##########  Phase 2 & 3  ##############
+def start_experiment_agent(hypothesis, plan):
+    if (not state_extract or not state_generate or not state_example):
+        return "", "", ""
+    global state_agent, step_index, state_complete
+    state_agent = True
+    step_index = 0
+    state_complete = False
+    # predefined_message = f"Implement the following hypothesis and experiment plan:\n\nHypothesis:\n{hypothesis}\n\nExperiment Plan:\n{plan}"
+    return example_data[index_ex]['code_init'], predefined_action_log, "", ""
+
+def submit_feedback(user_feedback, history, previous_response):
+    if (not state_extract or not  state_generate or not state_agent or not state_example):
+        return "", "", ""
+    global step_index, state_complete
+    step_index += 1
+    msg = history
+    if step_index < len(process_steps):
+        msg += previous_response + "\nUser feedback:" + user_feedback + "\n\n"
+        response_info = process_steps[step_index]
+        response = info_to_message(response_info)  # Convert dictionary to formatted string
+        response += "Please provide feedback based on the history, response entries, and observation, and questions: "
+        step_index += 1
+        msg += response
+    else:
+        state_complete = True
+        response = "Agent Finished."
+        
+    return msg, response, example_data[index_ex]['code_init'] if state_complete else example_data[index_ex]['code_final'], ""
+
+def load_phase_2_inputs(hypothesis, plan):
+    return hypothesis, plan, "# Code implementation will be displayed here after Start ExperimentAgent."
+
+
+
+predefined_action_log = """
+[Reasoning]: To understand the initial structure and functionality of train.py for effective improvements.
+[Action]: Inspect Script (train.py)
+Input: {"script_name": "train.py", "start_line_number": "1", "end_line_number": "74"}
+Objective: Understand the training script, including data processing, [...]
+[Observation]: The train.py script imports [...]. Sets random seeds [...]. Defines [...] Placeholder functions [...] exist without implementation. [...]
+[Feedback]: The script structure is clear, but key functions (train_model, predict) need proper implementation for proposed model training and prediction.\n
+"""
+
+
+predefined_observation = """
+Epoch [1/10],
+Train MSE: 0.543,
+Test MSE: 0.688
+Epoch [2/10],
+Train MSE: 0.242,
+Test MSE: 0.493\n
+"""
+
+# Initialize the global step_index and history
+process_steps = [
+    {
+        "Action": "Inspect Script Lines (train.py)",
+        "Observation": (
+            "The train.py script imports necessary libraries (e.g., pandas, sklearn, torch). "
+            "Sets random seeds for reproducibility. Defines compute_metrics_for_regression function "
+            "to calculate RMSE for different dimensions. Placeholder functions train_model and "
+            "predict exist without implementations."
+        ),
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The script executed successfully. Generated embeddings using the BERT model. Completed "
+            "the training process without errors. Metrics calculation placeholders indicated areas needing implementation."
+        ),
+    },
+    {
+        "Action": "Edit Script (train.py)",
+        "Observation": (
+            "Edited train.py to separate data loading, model definition, training loop, and evaluation into distinct functions. "
+            "The edited train.py now has clearly defined functions"
+            "for data loading (load_data), model definition (build_model), "
+            "training (train_model), and evaluation (evaluate_model). Similarly, eval.py is reorganized to load the model and perform predictions efficiently."
+        ),
+    },
+    {
+        "Action": "Retrieve Model",
+        "Observation": "CNN and BiLSTM retrieved.",
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The model trained over the specified number of epochs. Training and validation loss values are recorded for each epoch, "
+            "the decrease in loss indicates improved model performance."
+        )
+    },
+    {
+        "Action": "Evaluation",
+        "Observation": predefined_observation,
+    }
+]
+def info_to_message(info):
+    msg = ""
+    for k, v in info.items():
+        if isinstance(v, dict):
+            tempv = v
+            v = ""
+            for k2, v2 in tempv.items():
+                v += f"{k2}:\n  {v2}\n"
+        v = User.indent_text(v, 2)
+        msg += '-' * 64
+        msg += '\n'
+        msg += f"{k}:\n{v}\n"
+    return msg
+
+
+def handle_example_click(example_index):
+    global index_ex
+    index_ex = example_index
+    return load_example(index_ex)  # Simply return the text to display it in the textbox
+
+# Gradio Interface
+with gr.Blocks(css=".gr-examples-label {display: none;}", theme=gr.themes.Default()) as app:
+    gr.Markdown("# MLR- Copilot: Machine Learning Research based on LLM Agents")
+    gr.Markdown("###   ")
+    gr.Markdown("## <span style='color:Orange;'> This UI is for predefined example demo only.</span>")
+    gr.Markdown("## <span style='color:Orange;'> To reproduce the results please use [Github](https://github.com/du-nlp-lab/MLR-Copilot/)</span>")
+
+
+
+    gr.Markdown("MLR-Copilot is a framework where LLMs mimic researchers’ thought processes, designed to enhance the productivity of machine learning research by automating the generation and implementation of research ideas. It begins with a research paper, autonomously generating and validating these ideas, while incorporating human feedback to help reach executable research outcomes.")
+
+
+    gr.Markdown("##    ⬇️  Click an example at bottom to start！")
+
+
+    # Use state variables to store generated hypothesis and experiment plan
+    hypothesis_state = gr.State("")
+    experiment_plan_state = gr.State("")
+
+    ########## Phase 1:  Research Idea Generation Tab ##############
+    with gr.Tab("💡Stage 1: Research Idea Generation"):
+        gr.Markdown("### Extract Research Elements and Generate Research Ideas")
+        with gr.Row():
+            with gr.Column():
+                paper_text_input = gr.Textbox(value="", lines=10, label="📑 Research Paper Text", interactive=False)
+
+                # extract_button = gr.Button("🔍 Extract Research Elements")
+                # with gr.Row():
+                #     tasks_output = gr.Textbox(placeholder="Research task definition", label="Research Tasks", lines=2, interactive=False)
+                #     gaps_output = gr.Textbox(placeholder="Research gaps of current works", label="Research Gaps", lines=2, interactive=False)
+                #     keywords_output = gr.Textbox(placeholder="Paper keywords", label="Keywords", lines=2, interactive=False)
+                #     recent_works_output = gr.Textbox(placeholder="Recent works extracted from Semantic Scholar", label="Recent Works", lines=2, interactive=False)
+            with gr.Column():
+                with gr.Row():  # Move the button to the top 
+                    generate_button = gr.Button("✍️ Generate Research Hypothesis & Experiment Plan")
+                with gr.Group():
+                    gr.Markdown("### 🌟 Research Idea")
+                    with gr.Row():
+                        hypothesis_output = gr.Textbox(label="Generated Hypothesis", lines=20, interactive=False)
+                        experiment_plan_output = gr.Textbox(label="Generated Experiment Plan", lines=20, interactive=False)
+        
+    
+        # Step 1: Extract Research Elements
+        extract_button.click(
+            fn=extract_research_elements, 
+            inputs=paper_text_input, 
+            outputs=[tasks_output, gaps_output, keywords_output, recent_works_output]
+        )
+
+        generate_button.click(
+            fn=generate_and_store, 
+            inputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output], 
+            outputs=[hypothesis_output, experiment_plan_output, hypothesis_state, experiment_plan_state]
+        )
+
+        gr.Examples(
+            examples=example_text,
+            inputs=[paper_text_input],
+            outputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output, hypothesis_output, experiment_plan_output],
+            fn=load_example_and_set_index,
+            run_on_click = True,
+            # label="⬇️ Click an example to load"
+        )
+        
+
+
+
+    ########## Phase 2 & 3:  Experiment implementation and execution ##############
+    with gr.Tab("🧪 Stage 2 & Stage 3: Experiment implementation and execution"):
+        gr.Markdown("### Interact with the ExperimentAgent")
+
+        with gr.Row():
+            with gr.Column():
+                with gr.Group():
+                    gr.Markdown("### 🌟 Generated Research Idea")
+                    with gr.Row():
+                        idea_input = gr.Textbox(label="Generated Research Hypothesis", lines=30, interactive=False)
+                        plan_input = gr.Textbox(label="Generated Experiment Plan", lines=30, interactive=False)
+        
+            with gr.Column():
+                start_exp_agnet = gr.Button("⚙️ Start / Reset ExperimentAgent", elem_classes=["agent-btn"])
+                with gr.Group():
+                    gr.Markdown("### Implementation + Execution Log")    
+                    log = gr.Textbox(label="📖 Execution Log", lines=20, interactive=False)
+                    code_display = gr.Code(label="🧑‍💻 Implementation", language="python", interactive=False)
+
+            with gr.Column():
+                response = gr.Textbox(label="🤖 ExperimentAgent Response", lines=30, interactive=False)
+                feedback = gr.Textbox(placeholder="N/A", label="🧑‍🔬 User Feedback", lines=3, interactive=True)
+                submit_button = gr.Button("Submit", elem_classes=["Submit-btn"])
+        
+        hypothesis_state.change(
+            fn=load_phase_2_inputs, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[idea_input, plan_input, code_display]
+        )
+        
+        # Start research agent
+        start_exp_agnet.click(
+            fn=start_experiment_agent, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[code_display, log, response, feedback]
+        )
+        
+        submit_button.click(
+            fn=submit_feedback,
+            inputs=[feedback, log, response],
+            outputs=[log, response, code_display, feedback]
+        )
+
+# Test
+if __name__ == "__main__":
+    step_index = 0
+    app.launch(share=True)

.history/app_20250404171328.py

@@ -0,0 +1,329 @@
+import gradio as gr
+from pathlib import Path
+from reactagent.environment import Environment
+from reactagent.agents.agent_research import ResearchAgent
+from reactagent.runner import create_parser
+from reactagent import llm
+from reactagent.users.user import User
+import os
+import json
+
+
+# Global variables to store session state
+env = None
+agent = None
+state_example = False
+state_extract = False
+state_generate = False
+state_agent = False
+state_complete = False
+index_ex = "1"
+
+example_text = [
+    "Research Paper 1: Dataset and Baseline for Automatic Student Feedback Analysis",
+    "Research Paper 2: An Empirical Study on the Impact of Code Review on Software Quality"
+]
+
+# Load example JSON file
+def load_example_data():
+    with open("example/example_data.json", "r") as json_file:
+        example_data = json.load(json_file)
+    
+    for idx in example_data.keys():
+        try:
+            file = example_data[idx]["code_init"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_init"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")
+        try:
+            file = example_data[idx]["code_final"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_final"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")     
+    return example_data
+
+example_data = load_example_data()
+
+# Function to handle the selection of an example and populate the respective fields
+def load_example(example_id):
+    global index_ex 
+    index_ex = str(example_id)
+    example = example_data[index_ex]
+    paper_text = 'Title:\t' +  example['title'] + '\n\nAbstract:\t' + example['abstract']
+    return paper_text
+
+example_text = [load_example(1), load_example(2)]
+
+# Function to handle example clicks
+def load_example_and_set_index(paper_text_input):
+    global index_ex, state_example
+    state_example = True
+    index_ex = str(example_text.index(paper_text_input) + 1)
+    paper_text =  load_example(index_ex)
+
+    return paper_text,  "", "", "", "", "", ""
+
+
+
+########## Phase 1 ##############
+
+def extract_research_elements(paper_text):
+    global state_extract, index_ex, state_example
+    if not state_example or paper_text == "":
+        return "", "", "", ""
+    state_extract = True
+    if not paper_text.strip().startswith("Title:\t" + example_data[index_ex]["title"]):
+        print("Mismatch detected.")
+        print(paper_text)
+        return "", "", "", ""
+    example = example_data[index_ex]
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return tasks, gaps, keywords, recent_works
+
+
+# Step 2: Generate Research Hypothesis and Experiment Plan
+def generate_and_store(paper_text, tasks, gaps, keywords, recent_works):
+    if (not state_extract or not state_example or paper_text == ""):
+        return "", "", "", ""
+    global state_generate, index_ex
+    state_generate = True
+    hypothesis = example_data[index_ex]['hypothesis']
+    experiment_plan = example_data[index_ex]['experiment_plan']
+    return hypothesis, experiment_plan, hypothesis, experiment_plan
+
+##########  Phase 2 & 3  ##############
+def start_experiment_agent(hypothesis, plan):
+    if (not state_extract or not state_generate or not state_example):
+        return "", "", ""
+    global state_agent, step_index, state_complete
+    state_agent = True
+    step_index = 0
+    state_complete = False
+    # predefined_message = f"Implement the following hypothesis and experiment plan:\n\nHypothesis:\n{hypothesis}\n\nExperiment Plan:\n{plan}"
+    return example_data[index_ex]['code_init'], predefined_action_log, "", ""
+
+def submit_feedback(user_feedback, history, previous_response):
+    if (not state_extract or not  state_generate or not state_agent or not state_example):
+        return "", "", ""
+    global step_index, state_complete
+    step_index += 1
+    msg = history
+    if step_index < len(process_steps):
+        msg += previous_response + "\nUser feedback:" + user_feedback + "\n\n"
+        response_info = process_steps[step_index]
+        response = info_to_message(response_info)  # Convert dictionary to formatted string
+        response += "Please provide feedback based on the history, response entries, and observation, and questions: "
+        step_index += 1
+        msg += response
+    else:
+        state_complete = True
+        response = "Agent Finished."
+        
+    return msg, response, example_data[index_ex]['code_init'] if state_complete else example_data[index_ex]['code_final'], ""
+
+def load_phase_2_inputs(hypothesis, plan):
+    return hypothesis, plan, "# Code implementation will be displayed here after Start ExperimentAgent."
+
+
+
+predefined_action_log = """
+[Reasoning]: To understand the initial structure and functionality of train.py for effective improvements.
+[Action]: Inspect Script (train.py)
+Input: {"script_name": "train.py", "start_line_number": "1", "end_line_number": "74"}
+Objective: Understand the training script, including data processing, [...]
+[Observation]: The train.py script imports [...]. Sets random seeds [...]. Defines [...] Placeholder functions [...] exist without implementation. [...]
+[Feedback]: The script structure is clear, but key functions (train_model, predict) need proper implementation for proposed model training and prediction.\n
+"""
+
+
+predefined_observation = """
+Epoch [1/10],
+Train MSE: 0.543,
+Test MSE: 0.688
+Epoch [2/10],
+Train MSE: 0.242,
+Test MSE: 0.493\n
+"""
+
+# Initialize the global step_index and history
+process_steps = [
+    {
+        "Action": "Inspect Script Lines (train.py)",
+        "Observation": (
+            "The train.py script imports necessary libraries (e.g., pandas, sklearn, torch). "
+            "Sets random seeds for reproducibility. Defines compute_metrics_for_regression function "
+            "to calculate RMSE for different dimensions. Placeholder functions train_model and "
+            "predict exist without implementations."
+        ),
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The script executed successfully. Generated embeddings using the BERT model. Completed "
+            "the training process without errors. Metrics calculation placeholders indicated areas needing implementation."
+        ),
+    },
+    {
+        "Action": "Edit Script (train.py)",
+        "Observation": (
+            "Edited train.py to separate data loading, model definition, training loop, and evaluation into distinct functions. "
+            "The edited train.py now has clearly defined functions"
+            "for data loading (load_data), model definition (build_model), "
+            "training (train_model), and evaluation (evaluate_model). Similarly, eval.py is reorganized to load the model and perform predictions efficiently."
+        ),
+    },
+    {
+        "Action": "Retrieve Model",
+        "Observation": "CNN and BiLSTM retrieved.",
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The model trained over the specified number of epochs. Training and validation loss values are recorded for each epoch, "
+            "the decrease in loss indicates improved model performance."
+        )
+    },
+    {
+        "Action": "Evaluation",
+        "Observation": predefined_observation,
+    }
+]
+def info_to_message(info):
+    msg = ""
+    for k, v in info.items():
+        if isinstance(v, dict):
+            tempv = v
+            v = ""
+            for k2, v2 in tempv.items():
+                v += f"{k2}:\n  {v2}\n"
+        v = User.indent_text(v, 2)
+        msg += '-' * 64
+        msg += '\n'
+        msg += f"{k}:\n{v}\n"
+    return msg
+
+
+def handle_example_click(example_index):
+    global index_ex
+    index_ex = example_index
+    return load_example(index_ex)  # Simply return the text to display it in the textbox
+
+# Gradio Interface
+with gr.Blocks(css=".gr-examples-label {display: none;}", theme=gr.themes.Default()) as app:
+    gr.Markdown("# MLR- Copilot: Machine Learning Research based on LLM Agents")
+    gr.Markdown("###   ")
+    gr.Markdown("## <span style='color:Orange;'> This UI is for predefined example demo only.</span>")
+    gr.Markdown("## <span style='color:Orange;'> To reproduce the results please use [Github](https://github.com/du-nlp-lab/MLR-Copilot/)</span>")
+
+
+
+    gr.Markdown("MLR-Copilot is a framework where LLMs mimic researchers’ thought processes, designed to enhance the productivity of machine learning research by automating the generation and implementation of research ideas. It begins with a research paper, autonomously generating and validating these ideas, while incorporating human feedback to help reach executable research outcomes.")
+
+
+    gr.Markdown("##    ⬇️  Click an example at bottom to start！")
+
+
+    # Use state variables to store generated hypothesis and experiment plan
+    hypothesis_state = gr.State("")
+    experiment_plan_state = gr.State("")
+
+    ########## Phase 1:  Research Idea Generation Tab ##############
+    with gr.Tab("💡Stage 1: Research Idea Generation"):
+        gr.Markdown("### Extract Research Elements and Generate Research Ideas")
+        with gr.Row():
+            with gr.Column():
+                paper_text_input = gr.Textbox(value="", lines=10, label="📑 Research Paper Text", interactive=False)
+
+                extract_button = gr.Button("🔍 Extract Research Elements")
+                with gr.Row():
+                    tasks_output = gr.Textbox(placeholder="Research task definition", label="Research Tasks", lines=2, interactive=False)
+                    gaps_output = gr.Textbox(placeholder="Research gaps of current works", label="Research Gaps", lines=2, interactive=False)
+                    keywords_output = gr.Textbox(placeholder="Paper keywords", label="Keywords", lines=2, interactive=False)
+                    recent_works_output = gr.Textbox(placeholder="Recent works extracted from Semantic Scholar", label="Recent Works", lines=2, interactive=False)
+            with gr.Column():
+                with gr.Row():  # Move the button to the top 
+                    generate_button = gr.Button("✍️ Generate Research Hypothesis & Experiment Plan")
+                with gr.Group():
+                    gr.Markdown("### 🌟 Research Idea")
+                    with gr.Row():
+                        hypothesis_output = gr.Textbox(label="Generated Hypothesis", lines=20, interactive=False)
+                        experiment_plan_output = gr.Textbox(label="Generated Experiment Plan", lines=20, interactive=False)
+        
+    
+        # Step 1: Extract Research Elements
+        extract_button.click(
+            fn=extract_research_elements, 
+            inputs=paper_text_input, 
+            outputs=[tasks_output, gaps_output, keywords_output, recent_works_output]
+        )
+
+        generate_button.click(
+            fn=generate_and_store, 
+            inputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output], 
+            outputs=[hypothesis_output, experiment_plan_output, hypothesis_state, experiment_plan_state]
+        )
+
+        gr.Examples(
+            examples=example_text,
+            inputs=[paper_text_input],
+            outputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output, hypothesis_output, experiment_plan_output],
+            fn=load_example_and_set_index,
+            run_on_click = True,
+            # label="⬇️ Click an example to load"
+        )
+        
+
+
+
+    ########## Phase 2 & 3:  Experiment implementation and execution ##############
+    with gr.Tab("🧪 Stage 2 & Stage 3: Experiment implementation and execution"):
+        gr.Markdown("### Interact with the ExperimentAgent")
+
+        with gr.Row():
+            with gr.Column():
+                with gr.Group():
+                    gr.Markdown("### 🌟 Generated Research Idea")
+                    with gr.Row():
+                        idea_input = gr.Textbox(label="Generated Research Hypothesis", lines=30, interactive=False)
+                        plan_input = gr.Textbox(label="Generated Experiment Plan", lines=30, interactive=False)
+        
+            with gr.Column():
+                start_exp_agnet = gr.Button("⚙️ Start / Reset ExperimentAgent", elem_classes=["agent-btn"])
+                with gr.Group():
+                    gr.Markdown("### Implementation + Execution Log")    
+                    log = gr.Textbox(label="📖 Execution Log", lines=20, interactive=False)
+                    code_display = gr.Code(label="🧑‍💻 Implementation", language="python", interactive=False)
+
+            with gr.Column():
+                response = gr.Textbox(label="🤖 ExperimentAgent Response", lines=30, interactive=False)
+                feedback = gr.Textbox(placeholder="N/A", label="🧑‍🔬 User Feedback", lines=3, interactive=True)
+                submit_button = gr.Button("Submit", elem_classes=["Submit-btn"])
+        
+        hypothesis_state.change(
+            fn=load_phase_2_inputs, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[idea_input, plan_input, code_display]
+        )
+        
+        # Start research agent
+        start_exp_agnet.click(
+            fn=start_experiment_agent, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[code_display, log, response, feedback]
+        )
+        
+        submit_button.click(
+            fn=submit_feedback,
+            inputs=[feedback, log, response],
+            outputs=[log, response, code_display, feedback]
+        )
+
+# Test
+if __name__ == "__main__":
+    step_index = 0
+    app.launch(share=True)

.history/app_20250404171503.py

@@ -0,0 +1,329 @@
+import gradio as gr
+from pathlib import Path
+from reactagent.environment import Environment
+from reactagent.agents.agent_research import ResearchAgent
+from reactagent.runner import create_parser
+from reactagent import llm
+from reactagent.users.user import User
+import os
+import json
+
+
+# Global variables to store session state
+env = None
+agent = None
+state_example = False
+state_extract = False
+state_generate = False
+state_agent = False
+state_complete = False
+index_ex = "1"
+
+example_text = [
+    "Research Paper 1: Dataset and Baseline for Automatic Student Feedback Analysis",
+    "Research Paper 2: An Empirical Study on the Impact of Code Review on Software Quality"
+]
+
+# Load example JSON file
+def load_example_data():
+    with open("example/example_data.json", "r") as json_file:
+        example_data = json.load(json_file)
+    
+    for idx in example_data.keys():
+        try:
+            file = example_data[idx]["code_init"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_init"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")
+        try:
+            file = example_data[idx]["code_final"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_final"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")     
+    return example_data
+
+example_data = load_example_data()
+
+# Function to handle the selection of an example and populate the respective fields
+def load_example(example_id):
+    global index_ex 
+    index_ex = str(example_id)
+    example = example_data[index_ex]
+    paper_text = 'Title:\t' +  example['title'] + '\n\nAbstract:\t' + example['abstract']
+    return paper_text
+
+example_text = [load_example(1), load_example(2)]
+
+# Function to handle example clicks
+def load_example_and_set_index(paper_text_input):
+    global index_ex, state_example
+    state_example = True
+    index_ex = str(example_text.index(paper_text_input) + 1)
+    paper_text =  load_example(index_ex)
+
+    return paper_text,  "", "", "", "", "", ""
+
+
+
+########## Phase 1 ##############
+
+def extract_research_elements(paper_text):
+    global state_extract, index_ex, state_example
+    if not state_example or paper_text == "":
+        return "", "", "", ""
+    state_extract = True
+    if not paper_text.strip().startswith("Title:\t" + example_data[index_ex]["title"]):
+        print("Mismatch detected.")
+        print(paper_text)
+        return "", "", "", ""
+    example = example_data[index_ex]
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return tasks, gaps, keywords, recent_works
+
+
+# Step 2: Generate Research Hypothesis and Experiment Plan
+def generate_and_store(paper_text, tasks, gaps, keywords, recent_works):
+    if (not state_extract or not state_example or paper_text == ""):
+        return "", "", "", ""
+    global state_generate, index_ex
+    state_generate = True
+    hypothesis = example_data[index_ex]['hypothesis']
+    experiment_plan = example_data[index_ex]['experiment_plan']
+    return hypothesis, experiment_plan, hypothesis, experiment_plan
+
+##########  Phase 2 & 3  ##############
+def start_experiment_agent(hypothesis, plan):
+    if (not state_extract or not state_generate or not state_example):
+        return "", "", ""
+    global state_agent, step_index, state_complete
+    state_agent = True
+    step_index = 0
+    state_complete = False
+    # predefined_message = f"Implement the following hypothesis and experiment plan:\n\nHypothesis:\n{hypothesis}\n\nExperiment Plan:\n{plan}"
+    return example_data[index_ex]['code_init'], predefined_action_log, "", ""
+
+def submit_feedback(user_feedback, history, previous_response):
+    if (not state_extract or not  state_generate or not state_agent or not state_example):
+        return "", "", ""
+    global step_index, state_complete
+    step_index += 1
+    msg = history
+    if step_index < len(process_steps):
+        msg += previous_response + "\nUser feedback:" + user_feedback + "\n\n"
+        response_info = process_steps[step_index]
+        response = info_to_message(response_info)  # Convert dictionary to formatted string
+        response += "Please provide feedback based on the history, response entries, and observation, and questions: "
+        step_index += 1
+        msg += response
+    else:
+        state_complete = True
+        response = "Agent Finished."
+        
+    return msg, response, example_data[index_ex]['code_init'] if state_complete else example_data[index_ex]['code_final'], ""
+
+def load_phase_2_inputs(hypothesis, plan):
+    return hypothesis, plan, "# Code implementation will be displayed here after Start ExperimentAgent."
+
+
+
+predefined_action_log = """
+[Reasoning]: To understand the initial structure and functionality of train.py for effective improvements.
+[Action]: Inspect Script (train.py)
+Input: {"script_name": "train.py", "start_line_number": "1", "end_line_number": "74"}
+Objective: Understand the training script, including data processing, [...]
+[Observation]: The train.py script imports [...]. Sets random seeds [...]. Defines [...] Placeholder functions [...] exist without implementation. [...]
+[Feedback]: The script structure is clear, but key functions (train_model, predict) need proper implementation for proposed model training and prediction.\n
+"""
+
+
+predefined_observation = """
+Epoch [1/10],
+Train MSE: 0.543,
+Test MSE: 0.688
+Epoch [2/10],
+Train MSE: 0.242,
+Test MSE: 0.493\n
+"""
+
+# Initialize the global step_index and history
+process_steps = [
+    {
+        "Action": "Inspect Script Lines (train.py)",
+        "Observation": (
+            "The train.py script imports necessary libraries (e.g., pandas, sklearn, torch). "
+            "Sets random seeds for reproducibility. Defines compute_metrics_for_regression function "
+            "to calculate RMSE for different dimensions. Placeholder functions train_model and "
+            "predict exist without implementations."
+        ),
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The script executed successfully. Generated embeddings using the BERT model. Completed "
+            "the training process without errors. Metrics calculation placeholders indicated areas needing implementation."
+        ),
+    },
+    {
+        "Action": "Edit Script (train.py)",
+        "Observation": (
+            "Edited train.py to separate data loading, model definition, training loop, and evaluation into distinct functions. "
+            "The edited train.py now has clearly defined functions"
+            "for data loading (load_data), model definition (build_model), "
+            "training (train_model), and evaluation (evaluate_model). Similarly, eval.py is reorganized to load the model and perform predictions efficiently."
+        ),
+    },
+    {
+        "Action": "Retrieve Model",
+        "Observation": "CNN and BiLSTM retrieved.",
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The model trained over the specified number of epochs. Training and validation loss values are recorded for each epoch, "
+            "the decrease in loss indicates improved model performance."
+        )
+    },
+    {
+        "Action": "Evaluation",
+        "Observation": predefined_observation,
+    }
+]
+def info_to_message(info):
+    msg = ""
+    for k, v in info.items():
+        if isinstance(v, dict):
+            tempv = v
+            v = ""
+            for k2, v2 in tempv.items():
+                v += f"{k2}:\n  {v2}\n"
+        v = User.indent_text(v, 2)
+        msg += '-' * 64
+        msg += '\n'
+        msg += f"{k}:\n{v}\n"
+    return msg
+
+
+def handle_example_click(example_index):
+    global index_ex
+    index_ex = example_index
+    return load_example(index_ex)  # Simply return the text to display it in the textbox
+
+# Gradio Interface
+with gr.Blocks(css=".gr-examples-label {display: none;}", theme=gr.themes.Default()) as app:
+    gr.Markdown("# MLR- Copilot: Machine Learning Research based on LLM Agents")
+    gr.Markdown("###   ")
+    gr.Markdown("## <span style='color:Orange;'> This UI is for predefined example demo only.</span>")
+    gr.Markdown("## <span style='color:Orange;'> To reproduce the results please use [Github](https://github.com/du-nlp-lab/MLR-Copilot/)</span>")
+
+
+
+    gr.Markdown("MLR-Copilot is a framework where LLMs mimic researchers’ thought processes, designed to enhance the productivity of machine learning research by automating the generation and implementation of research ideas. It begins with a research paper, autonomously generating and validating these ideas, while incorporating human feedback to help reach executable research outcomes.")
+
+
+    gr.Markdown("##    ⬇️  Click an example at bottom to start！")
+
+
+    # Use state variables to store generated hypothesis and experiment plan
+    hypothesis_state = gr.State("")
+    experiment_plan_state = gr.State("")
+
+    ########## Phase 1:  Research Idea Generation Tab ##############
+    with gr.Tab("💡Stage 1: Research Idea Generation"):
+        gr.Markdown("### Extract Research Elements and Generate Research Ideas")
+        with gr.Row():
+            with gr.Column():
+                paper_text_input = gr.Textbox(value="", lines=10, label="📑 Research Paper Text", interactive=False)
+
+                # extract_button = gr.Button("🔍 Extract Research Elements")
+                with gr.Row():
+                    tasks_output = gr.Textbox(placeholder="Research task definition", label="Research Tasks", lines=2, interactive=False)
+                    gaps_output = gr.Textbox(placeholder="Research gaps of current works", label="Research Gaps", lines=2, interactive=False)
+                    keywords_output = gr.Textbox(placeholder="Paper keywords", label="Keywords", lines=2, interactive=False)
+                    recent_works_output = gr.Textbox(placeholder="Recent works extracted from Semantic Scholar", label="Recent Works", lines=2, interactive=False)
+            with gr.Column():
+                with gr.Row():  # Move the button to the top 
+                    generate_button = gr.Button("✍️ Generate Research Hypothesis & Experiment Plan")
+                with gr.Group():
+                    gr.Markdown("### 🌟 Research Idea")
+                    with gr.Row():
+                        hypothesis_output = gr.Textbox(label="Generated Hypothesis", lines=20, interactive=False)
+                        experiment_plan_output = gr.Textbox(label="Generated Experiment Plan", lines=20, interactive=False)
+        
+    
+        # Step 1: Extract Research Elements
+        extract_button.click(
+            fn=extract_research_elements, 
+            inputs=paper_text_input, 
+            outputs=[tasks_output, gaps_output, keywords_output, recent_works_output]
+        )
+
+        generate_button.click(
+            fn=generate_and_store, 
+            inputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output], 
+            outputs=[hypothesis_output, experiment_plan_output, hypothesis_state, experiment_plan_state]
+        )
+
+        gr.Examples(
+            examples=example_text,
+            inputs=[paper_text_input],
+            outputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output, hypothesis_output, experiment_plan_output],
+            fn=load_example_and_set_index,
+            run_on_click = True,
+            # label="⬇️ Click an example to load"
+        )
+        
+
+
+
+    ########## Phase 2 & 3:  Experiment implementation and execution ##############
+    with gr.Tab("🧪 Stage 2 & Stage 3: Experiment implementation and execution"):
+        gr.Markdown("### Interact with the ExperimentAgent")
+
+        with gr.Row():
+            with gr.Column():
+                with gr.Group():
+                    gr.Markdown("### 🌟 Generated Research Idea")
+                    with gr.Row():
+                        idea_input = gr.Textbox(label="Generated Research Hypothesis", lines=30, interactive=False)
+                        plan_input = gr.Textbox(label="Generated Experiment Plan", lines=30, interactive=False)
+        
+            with gr.Column():
+                start_exp_agnet = gr.Button("⚙️ Start / Reset ExperimentAgent", elem_classes=["agent-btn"])
+                with gr.Group():
+                    gr.Markdown("### Implementation + Execution Log")    
+                    log = gr.Textbox(label="📖 Execution Log", lines=20, interactive=False)
+                    code_display = gr.Code(label="🧑‍💻 Implementation", language="python", interactive=False)
+
+            with gr.Column():
+                response = gr.Textbox(label="🤖 ExperimentAgent Response", lines=30, interactive=False)
+                feedback = gr.Textbox(placeholder="N/A", label="🧑‍🔬 User Feedback", lines=3, interactive=True)
+                submit_button = gr.Button("Submit", elem_classes=["Submit-btn"])
+        
+        hypothesis_state.change(
+            fn=load_phase_2_inputs, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[idea_input, plan_input, code_display]
+        )
+        
+        # Start research agent
+        start_exp_agnet.click(
+            fn=start_experiment_agent, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[code_display, log, response, feedback]
+        )
+        
+        submit_button.click(
+            fn=submit_feedback,
+            inputs=[feedback, log, response],
+            outputs=[log, response, code_display, feedback]
+        )
+
+# Test
+if __name__ == "__main__":
+    step_index = 0
+    app.launch(share=True)

.history/app_20250404171504.py

@@ -0,0 +1,329 @@
+import gradio as gr
+from pathlib import Path
+from reactagent.environment import Environment
+from reactagent.agents.agent_research import ResearchAgent
+from reactagent.runner import create_parser
+from reactagent import llm
+from reactagent.users.user import User
+import os
+import json
+
+
+# Global variables to store session state
+env = None
+agent = None
+state_example = False
+state_extract = False
+state_generate = False
+state_agent = False
+state_complete = False
+index_ex = "1"
+
+example_text = [
+    "Research Paper 1: Dataset and Baseline for Automatic Student Feedback Analysis",
+    "Research Paper 2: An Empirical Study on the Impact of Code Review on Software Quality"
+]
+
+# Load example JSON file
+def load_example_data():
+    with open("example/example_data.json", "r") as json_file:
+        example_data = json.load(json_file)
+    
+    for idx in example_data.keys():
+        try:
+            file = example_data[idx]["code_init"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_init"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")
+        try:
+            file = example_data[idx]["code_final"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_final"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")     
+    return example_data
+
+example_data = load_example_data()
+
+# Function to handle the selection of an example and populate the respective fields
+def load_example(example_id):
+    global index_ex 
+    index_ex = str(example_id)
+    example = example_data[index_ex]
+    paper_text = 'Title:\t' +  example['title'] + '\n\nAbstract:\t' + example['abstract']
+    return paper_text
+
+example_text = [load_example(1), load_example(2)]
+
+# Function to handle example clicks
+def load_example_and_set_index(paper_text_input):
+    global index_ex, state_example
+    state_example = True
+    index_ex = str(example_text.index(paper_text_input) + 1)
+    paper_text =  load_example(index_ex)
+
+    return paper_text,  "", "", "", "", "", ""
+
+
+
+########## Phase 1 ##############
+
+def extract_research_elements(paper_text):
+    global state_extract, index_ex, state_example
+    if not state_example or paper_text == "":
+        return "", "", "", ""
+    state_extract = True
+    if not paper_text.strip().startswith("Title:\t" + example_data[index_ex]["title"]):
+        print("Mismatch detected.")
+        print(paper_text)
+        return "", "", "", ""
+    example = example_data[index_ex]
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return tasks, gaps, keywords, recent_works
+
+
+# Step 2: Generate Research Hypothesis and Experiment Plan
+def generate_and_store(paper_text, tasks, gaps, keywords, recent_works):
+    if (not state_extract or not state_example or paper_text == ""):
+        return "", "", "", ""
+    global state_generate, index_ex
+    state_generate = True
+    hypothesis = example_data[index_ex]['hypothesis']
+    experiment_plan = example_data[index_ex]['experiment_plan']
+    return hypothesis, experiment_plan, hypothesis, experiment_plan
+
+##########  Phase 2 & 3  ##############
+def start_experiment_agent(hypothesis, plan):
+    if (not state_extract or not state_generate or not state_example):
+        return "", "", ""
+    global state_agent, step_index, state_complete
+    state_agent = True
+    step_index = 0
+    state_complete = False
+    # predefined_message = f"Implement the following hypothesis and experiment plan:\n\nHypothesis:\n{hypothesis}\n\nExperiment Plan:\n{plan}"
+    return example_data[index_ex]['code_init'], predefined_action_log, "", ""
+
+def submit_feedback(user_feedback, history, previous_response):
+    if (not state_extract or not  state_generate or not state_agent or not state_example):
+        return "", "", ""
+    global step_index, state_complete
+    step_index += 1
+    msg = history
+    if step_index < len(process_steps):
+        msg += previous_response + "\nUser feedback:" + user_feedback + "\n\n"
+        response_info = process_steps[step_index]
+        response = info_to_message(response_info)  # Convert dictionary to formatted string
+        response += "Please provide feedback based on the history, response entries, and observation, and questions: "
+        step_index += 1
+        msg += response
+    else:
+        state_complete = True
+        response = "Agent Finished."
+        
+    return msg, response, example_data[index_ex]['code_init'] if state_complete else example_data[index_ex]['code_final'], ""
+
+def load_phase_2_inputs(hypothesis, plan):
+    return hypothesis, plan, "# Code implementation will be displayed here after Start ExperimentAgent."
+
+
+
+predefined_action_log = """
+[Reasoning]: To understand the initial structure and functionality of train.py for effective improvements.
+[Action]: Inspect Script (train.py)
+Input: {"script_name": "train.py", "start_line_number": "1", "end_line_number": "74"}
+Objective: Understand the training script, including data processing, [...]
+[Observation]: The train.py script imports [...]. Sets random seeds [...]. Defines [...] Placeholder functions [...] exist without implementation. [...]
+[Feedback]: The script structure is clear, but key functions (train_model, predict) need proper implementation for proposed model training and prediction.\n
+"""
+
+
+predefined_observation = """
+Epoch [1/10],
+Train MSE: 0.543,
+Test MSE: 0.688
+Epoch [2/10],
+Train MSE: 0.242,
+Test MSE: 0.493\n
+"""
+
+# Initialize the global step_index and history
+process_steps = [
+    {
+        "Action": "Inspect Script Lines (train.py)",
+        "Observation": (
+            "The train.py script imports necessary libraries (e.g., pandas, sklearn, torch). "
+            "Sets random seeds for reproducibility. Defines compute_metrics_for_regression function "
+            "to calculate RMSE for different dimensions. Placeholder functions train_model and "
+            "predict exist without implementations."
+        ),
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The script executed successfully. Generated embeddings using the BERT model. Completed "
+            "the training process without errors. Metrics calculation placeholders indicated areas needing implementation."
+        ),
+    },
+    {
+        "Action": "Edit Script (train.py)",
+        "Observation": (
+            "Edited train.py to separate data loading, model definition, training loop, and evaluation into distinct functions. "
+            "The edited train.py now has clearly defined functions"
+            "for data loading (load_data), model definition (build_model), "
+            "training (train_model), and evaluation (evaluate_model). Similarly, eval.py is reorganized to load the model and perform predictions efficiently."
+        ),
+    },
+    {
+        "Action": "Retrieve Model",
+        "Observation": "CNN and BiLSTM retrieved.",
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The model trained over the specified number of epochs. Training and validation loss values are recorded for each epoch, "
+            "the decrease in loss indicates improved model performance."
+        )
+    },
+    {
+        "Action": "Evaluation",
+        "Observation": predefined_observation,
+    }
+]
+def info_to_message(info):
+    msg = ""
+    for k, v in info.items():
+        if isinstance(v, dict):
+            tempv = v
+            v = ""
+            for k2, v2 in tempv.items():
+                v += f"{k2}:\n  {v2}\n"
+        v = User.indent_text(v, 2)
+        msg += '-' * 64
+        msg += '\n'
+        msg += f"{k}:\n{v}\n"
+    return msg
+
+
+def handle_example_click(example_index):
+    global index_ex
+    index_ex = example_index
+    return load_example(index_ex)  # Simply return the text to display it in the textbox
+
+# Gradio Interface
+with gr.Blocks(css=".gr-examples-label {display: none;}", theme=gr.themes.Default()) as app:
+    gr.Markdown("# MLR- Copilot: Machine Learning Research based on LLM Agents")
+    gr.Markdown("###   ")
+    gr.Markdown("## <span style='color:Orange;'> This UI is for predefined example demo only.</span>")
+    gr.Markdown("## <span style='color:Orange;'> To reproduce the results please use [Github](https://github.com/du-nlp-lab/MLR-Copilot/)</span>")
+
+
+
+    gr.Markdown("MLR-Copilot is a framework where LLMs mimic researchers’ thought processes, designed to enhance the productivity of machine learning research by automating the generation and implementation of research ideas. It begins with a research paper, autonomously generating and validating these ideas, while incorporating human feedback to help reach executable research outcomes.")
+
+
+    gr.Markdown("##    ⬇️  Click an example at bottom to start！")
+
+
+    # Use state variables to store generated hypothesis and experiment plan
+    hypothesis_state = gr.State("")
+    experiment_plan_state = gr.State("")
+
+    ########## Phase 1:  Research Idea Generation Tab ##############
+    with gr.Tab("💡Stage 1: Research Idea Generation"):
+        gr.Markdown("### Extract Research Elements and Generate Research Ideas")
+        with gr.Row():
+            with gr.Column():
+                paper_text_input = gr.Textbox(value="", lines=10, label="📑 Research Paper Text", interactive=False)
+
+                # extract_button = gr.Button("🔍 Extract Research Elements")
+                with gr.Row():
+                    tasks_output = gr.Textbox(placeholder="Research task definition", label="Research Tasks", lines=2, interactive=False)
+                    gaps_output = gr.Textbox(placeholder="Research gaps of current works", label="Research Gaps", lines=2, interactive=False)
+                    keywords_output = gr.Textbox(placeholder="Paper keywords", label="Keywords", lines=2, interactive=False)
+                    recent_works_output = gr.Textbox(placeholder="Recent works extracted from Semantic Scholar", label="Recent Works", lines=2, interactive=False)
+            with gr.Column():
+                with gr.Row():  # Move the button to the top 
+                    generate_button = gr.Button("✍️ Generate Research Hypothesis & Experiment Plan")
+                with gr.Group():
+                    gr.Markdown("### 🌟 Research Idea")
+                    with gr.Row():
+                        hypothesis_output = gr.Textbox(label="Generated Hypothesis", lines=20, interactive=False)
+                        experiment_plan_output = gr.Textbox(label="Generated Experiment Plan", lines=20, interactive=False)
+        
+    
+        # Step 1: Extract Research Elements
+        extract_button.click(
+            fn=extract_research_elements, 
+            inputs=paper_text_input, 
+            outputs=[tasks_output, gaps_output, keywords_output, recent_works_output]
+        )
+
+        generate_button.click(
+            fn=generate_and_store, 
+            inputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output], 
+            outputs=[hypothesis_output, experiment_plan_output, hypothesis_state, experiment_plan_state]
+        )
+
+        gr.Examples(
+            examples=example_text,
+            inputs=[paper_text_input],
+            outputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output, hypothesis_output, experiment_plan_output],
+            fn=load_example_and_set_index,
+            run_on_click = True,
+            # label="⬇️ Click an example to load"
+        )
+        
+
+
+
+    ########## Phase 2 & 3:  Experiment implementation and execution ##############
+    with gr.Tab("🧪 Stage 2 & Stage 3: Experiment implementation and execution"):
+        gr.Markdown("### Interact with the ExperimentAgent")
+
+        with gr.Row():
+            with gr.Column():
+                with gr.Group():
+                    gr.Markdown("### 🌟 Generated Research Idea")
+                    with gr.Row():
+                        idea_input = gr.Textbox(label="Generated Research Hypothesis", lines=30, interactive=False)
+                        plan_input = gr.Textbox(label="Generated Experiment Plan", lines=30, interactive=False)
+        
+            with gr.Column():
+                start_exp_agnet = gr.Button("⚙️ Start / Reset ExperimentAgent", elem_classes=["agent-btn"])
+                with gr.Group():
+                    gr.Markdown("### Implementation + Execution Log")    
+                    log = gr.Textbox(label="📖 Execution Log", lines=20, interactive=False)
+                    code_display = gr.Code(label="🧑‍💻 Implementation", language="python", interactive=False)
+
+            with gr.Column():
+                response = gr.Textbox(label="🤖 ExperimentAgent Response", lines=30, interactive=False)
+                feedback = gr.Textbox(placeholder="N/A", label="🧑‍🔬 User Feedback", lines=3, interactive=True)
+                submit_button = gr.Button("Submit", elem_classes=["Submit-btn"])
+        
+        hypothesis_state.change(
+            fn=load_phase_2_inputs, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[idea_input, plan_input, code_display]
+        )
+        
+        # Start research agent
+        start_exp_agnet.click(
+            fn=start_experiment_agent, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[code_display, log, response, feedback]
+        )
+        
+        submit_button.click(
+            fn=submit_feedback,
+            inputs=[feedback, log, response],
+            outputs=[log, response, code_display, feedback]
+        )
+
+# Test
+if __name__ == "__main__":
+    step_index = 0
+    app.launch(share=True)

.history/app_20250404171505.py

@@ -0,0 +1,329 @@
+import gradio as gr
+from pathlib import Path
+from reactagent.environment import Environment
+from reactagent.agents.agent_research import ResearchAgent
+from reactagent.runner import create_parser
+from reactagent import llm
+from reactagent.users.user import User
+import os
+import json
+
+
+# Global variables to store session state
+env = None
+agent = None
+state_example = False
+state_extract = False
+state_generate = False
+state_agent = False
+state_complete = False
+index_ex = "1"
+
+example_text = [
+    "Research Paper 1: Dataset and Baseline for Automatic Student Feedback Analysis",
+    "Research Paper 2: An Empirical Study on the Impact of Code Review on Software Quality"
+]
+
+# Load example JSON file
+def load_example_data():
+    with open("example/example_data.json", "r") as json_file:
+        example_data = json.load(json_file)
+    
+    for idx in example_data.keys():
+        try:
+            file = example_data[idx]["code_init"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_init"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")
+        try:
+            file = example_data[idx]["code_final"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_final"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")     
+    return example_data
+
+example_data = load_example_data()
+
+# Function to handle the selection of an example and populate the respective fields
+def load_example(example_id):
+    global index_ex 
+    index_ex = str(example_id)
+    example = example_data[index_ex]
+    paper_text = 'Title:\t' +  example['title'] + '\n\nAbstract:\t' + example['abstract']
+    return paper_text
+
+example_text = [load_example(1), load_example(2)]
+
+# Function to handle example clicks
+def load_example_and_set_index(paper_text_input):
+    global index_ex, state_example
+    state_example = True
+    index_ex = str(example_text.index(paper_text_input) + 1)
+    paper_text =  load_example(index_ex)
+
+    return paper_text,  "", "", "", "", "", ""
+
+
+
+########## Phase 1 ##############
+
+def extract_research_elements(paper_text):
+    global state_extract, index_ex, state_example
+    if not state_example or paper_text == "":
+        return "", "", "", ""
+    state_extract = True
+    if not paper_text.strip().startswith("Title:\t" + example_data[index_ex]["title"]):
+        print("Mismatch detected.")
+        print(paper_text)
+        return "", "", "", ""
+    example = example_data[index_ex]
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return tasks, gaps, keywords, recent_works
+
+
+# Step 2: Generate Research Hypothesis and Experiment Plan
+def generate_and_store(paper_text, tasks, gaps, keywords, recent_works):
+    if (not state_extract or not state_example or paper_text == ""):
+        return "", "", "", ""
+    global state_generate, index_ex
+    state_generate = True
+    hypothesis = example_data[index_ex]['hypothesis']
+    experiment_plan = example_data[index_ex]['experiment_plan']
+    return hypothesis, experiment_plan, hypothesis, experiment_plan
+
+##########  Phase 2 & 3  ##############
+def start_experiment_agent(hypothesis, plan):
+    if (not state_extract or not state_generate or not state_example):
+        return "", "", ""
+    global state_agent, step_index, state_complete
+    state_agent = True
+    step_index = 0
+    state_complete = False
+    # predefined_message = f"Implement the following hypothesis and experiment plan:\n\nHypothesis:\n{hypothesis}\n\nExperiment Plan:\n{plan}"
+    return example_data[index_ex]['code_init'], predefined_action_log, "", ""
+
+def submit_feedback(user_feedback, history, previous_response):
+    if (not state_extract or not  state_generate or not state_agent or not state_example):
+        return "", "", ""
+    global step_index, state_complete
+    step_index += 1
+    msg = history
+    if step_index < len(process_steps):
+        msg += previous_response + "\nUser feedback:" + user_feedback + "\n\n"
+        response_info = process_steps[step_index]
+        response = info_to_message(response_info)  # Convert dictionary to formatted string
+        response += "Please provide feedback based on the history, response entries, and observation, and questions: "
+        step_index += 1
+        msg += response
+    else:
+        state_complete = True
+        response = "Agent Finished."
+        
+    return msg, response, example_data[index_ex]['code_init'] if state_complete else example_data[index_ex]['code_final'], ""
+
+def load_phase_2_inputs(hypothesis, plan):
+    return hypothesis, plan, "# Code implementation will be displayed here after Start ExperimentAgent."
+
+
+
+predefined_action_log = """
+[Reasoning]: To understand the initial structure and functionality of train.py for effective improvements.
+[Action]: Inspect Script (train.py)
+Input: {"script_name": "train.py", "start_line_number": "1", "end_line_number": "74"}
+Objective: Understand the training script, including data processing, [...]
+[Observation]: The train.py script imports [...]. Sets random seeds [...]. Defines [...] Placeholder functions [...] exist without implementation. [...]
+[Feedback]: The script structure is clear, but key functions (train_model, predict) need proper implementation for proposed model training and prediction.\n
+"""
+
+
+predefined_observation = """
+Epoch [1/10],
+Train MSE: 0.543,
+Test MSE: 0.688
+Epoch [2/10],
+Train MSE: 0.242,
+Test MSE: 0.493\n
+"""
+
+# Initialize the global step_index and history
+process_steps = [
+    {
+        "Action": "Inspect Script Lines (train.py)",
+        "Observation": (
+            "The train.py script imports necessary libraries (e.g., pandas, sklearn, torch). "
+            "Sets random seeds for reproducibility. Defines compute_metrics_for_regression function "
+            "to calculate RMSE for different dimensions. Placeholder functions train_model and "
+            "predict exist without implementations."
+        ),
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The script executed successfully. Generated embeddings using the BERT model. Completed "
+            "the training process without errors. Metrics calculation placeholders indicated areas needing implementation."
+        ),
+    },
+    {
+        "Action": "Edit Script (train.py)",
+        "Observation": (
+            "Edited train.py to separate data loading, model definition, training loop, and evaluation into distinct functions. "
+            "The edited train.py now has clearly defined functions"
+            "for data loading (load_data), model definition (build_model), "
+            "training (train_model), and evaluation (evaluate_model). Similarly, eval.py is reorganized to load the model and perform predictions efficiently."
+        ),
+    },
+    {
+        "Action": "Retrieve Model",
+        "Observation": "CNN and BiLSTM retrieved.",
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The model trained over the specified number of epochs. Training and validation loss values are recorded for each epoch, "
+            "the decrease in loss indicates improved model performance."
+        )
+    },
+    {
+        "Action": "Evaluation",
+        "Observation": predefined_observation,
+    }
+]
+def info_to_message(info):
+    msg = ""
+    for k, v in info.items():
+        if isinstance(v, dict):
+            tempv = v
+            v = ""
+            for k2, v2 in tempv.items():
+                v += f"{k2}:\n  {v2}\n"
+        v = User.indent_text(v, 2)
+        msg += '-' * 64
+        msg += '\n'
+        msg += f"{k}:\n{v}\n"
+    return msg
+
+
+def handle_example_click(example_index):
+    global index_ex
+    index_ex = example_index
+    return load_example(index_ex)  # Simply return the text to display it in the textbox
+
+# Gradio Interface
+with gr.Blocks(css=".gr-examples-label {display: none;}", theme=gr.themes.Default()) as app:
+    gr.Markdown("# MLR- Copilot: Machine Learning Research based on LLM Agents")
+    gr.Markdown("###   ")
+    gr.Markdown("## <span style='color:Orange;'> This UI is for predefined example demo only.</span>")
+    gr.Markdown("## <span style='color:Orange;'> To reproduce the results please use [Github](https://github.com/du-nlp-lab/MLR-Copilot/)</span>")
+
+
+
+    gr.Markdown("MLR-Copilot is a framework where LLMs mimic researchers’ thought processes, designed to enhance the productivity of machine learning research by automating the generation and implementation of research ideas. It begins with a research paper, autonomously generating and validating these ideas, while incorporating human feedback to help reach executable research outcomes.")
+
+
+    gr.Markdown("##    ⬇️  Click an example at bottom to start！")
+
+
+    # Use state variables to store generated hypothesis and experiment plan
+    hypothesis_state = gr.State("")
+    experiment_plan_state = gr.State("")
+
+    ########## Phase 1:  Research Idea Generation Tab ##############
+    with gr.Tab("💡Stage 1: Research Idea Generation"):
+        gr.Markdown("### Extract Research Elements and Generate Research Ideas")
+        with gr.Row():
+            with gr.Column():
+                paper_text_input = gr.Textbox(value="", lines=10, label="📑 Research Paper Text", interactive=False)
+
+                # extract_button = gr.Button("🔍 Extract Research Elements")
+                with gr.Row():
+                    tasks_output = gr.Textbox(placeholder="Research task definition", label="Research Tasks", lines=2, interactive=False)
+                    gaps_output = gr.Textbox(placeholder="Research gaps of current works", label="Research Gaps", lines=2, interactive=False)
+                    keywords_output = gr.Textbox(placeholder="Paper keywords", label="Keywords", lines=2, interactive=False)
+                    recent_works_output = gr.Textbox(placeholder="Recent works extracted from Semantic Scholar", label="Recent Works", lines=2, interactive=False)
+            with gr.Column():
+                with gr.Row():  # Move the button to the top 
+                    generate_button = gr.Button("✍️ Generate Research Hypothesis & Experiment Plan")
+                with gr.Group():
+                    gr.Markdown("### 🌟 Research Idea")
+                    with gr.Row():
+                        hypothesis_output = gr.Textbox(label="Generated Hypothesis", lines=20, interactive=False)
+                        experiment_plan_output = gr.Textbox(label="Generated Experiment Plan", lines=20, interactive=False)
+        
+    
+        # Step 1: Extract Research Elements
+        extract_button.click(
+            fn=extract_research_elements, 
+            inputs=paper_text_input, 
+            outputs=[tasks_output, gaps_output, keywords_output, recent_works_output]
+        )
+
+        generate_button.click(
+            fn=generate_and_store, 
+            inputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output], 
+            outputs=[hypothesis_output, experiment_plan_output, hypothesis_state, experiment_plan_state]
+        )
+
+        gr.Examples(
+            examples=example_text,
+            inputs=[paper_text_input],
+            outputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output, hypothesis_output, experiment_plan_output],
+            fn=load_example_and_set_index,
+            run_on_click = True,
+            # label="⬇️ Click an example to load"
+        )
+        
+
+
+
+    ########## Phase 2 & 3:  Experiment implementation and execution ##############
+    with gr.Tab("🧪 Stage 2 & Stage 3: Experiment implementation and execution"):
+        gr.Markdown("### Interact with the ExperimentAgent")
+
+        with gr.Row():
+            with gr.Column():
+                with gr.Group():
+                    gr.Markdown("### 🌟 Generated Research Idea")
+                    with gr.Row():
+                        idea_input = gr.Textbox(label="Generated Research Hypothesis", lines=30, interactive=False)
+                        plan_input = gr.Textbox(label="Generated Experiment Plan", lines=30, interactive=False)
+        
+            with gr.Column():
+                start_exp_agnet = gr.Button("⚙️ Start / Reset ExperimentAgent", elem_classes=["agent-btn"])
+                with gr.Group():
+                    gr.Markdown("### Implementation + Execution Log")    
+                    log = gr.Textbox(label="📖 Execution Log", lines=20, interactive=False)
+                    code_display = gr.Code(label="🧑‍💻 Implementation", language="python", interactive=False)
+
+            with gr.Column():
+                response = gr.Textbox(label="🤖 ExperimentAgent Response", lines=30, interactive=False)
+                feedback = gr.Textbox(placeholder="N/A", label="🧑‍🔬 User Feedback", lines=3, interactive=True)
+                submit_button = gr.Button("Submit", elem_classes=["Submit-btn"])
+        
+        hypothesis_state.change(
+            fn=load_phase_2_inputs, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[idea_input, plan_input, code_display]
+        )
+        
+        # Start research agent
+        start_exp_agnet.click(
+            fn=start_experiment_agent, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[code_display, log, response, feedback]
+        )
+        
+        submit_button.click(
+            fn=submit_feedback,
+            inputs=[feedback, log, response],
+            outputs=[log, response, code_display, feedback]
+        )
+
+# Test
+if __name__ == "__main__":
+    step_index = 0
+    app.launch(share=True)

.history/app_20250404171523.py

@@ -0,0 +1,329 @@
+import gradio as gr
+from pathlib import Path
+from reactagent.environment import Environment
+from reactagent.agents.agent_research import ResearchAgent
+from reactagent.runner import create_parser
+from reactagent import llm
+from reactagent.users.user import User
+import os
+import json
+
+
+# Global variables to store session state
+env = None
+agent = None
+state_example = False
+state_extract = False
+state_generate = False
+state_agent = False
+state_complete = False
+index_ex = "1"
+
+example_text = [
+    "Research Paper 1: Dataset and Baseline for Automatic Student Feedback Analysis",
+    "Research Paper 2: An Empirical Study on the Impact of Code Review on Software Quality"
+]
+
+# Load example JSON file
+def load_example_data():
+    with open("example/example_data.json", "r") as json_file:
+        example_data = json.load(json_file)
+    
+    for idx in example_data.keys():
+        try:
+            file = example_data[idx]["code_init"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_init"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")
+        try:
+            file = example_data[idx]["code_final"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_final"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")     
+    return example_data
+
+example_data = load_example_data()
+
+# Function to handle the selection of an example and populate the respective fields
+def load_example(example_id):
+    global index_ex 
+    index_ex = str(example_id)
+    example = example_data[index_ex]
+    paper_text = 'Title:\t' +  example['title'] + '\n\nAbstract:\t' + example['abstract']
+    return paper_text
+
+example_text = [load_example(1), load_example(2)]
+
+# Function to handle example clicks
+def load_example_and_set_index(paper_text_input):
+    global index_ex, state_example
+    state_example = True
+    index_ex = str(example_text.index(paper_text_input) + 1)
+    paper_text =  load_example(index_ex)
+
+    return paper_text,  "", "", "", "", "", ""
+
+
+
+########## Phase 1 ##############
+
+def extract_research_elements(paper_text):
+    global state_extract, index_ex, state_example
+    if not state_example or paper_text == "":
+        return "", "", "", ""
+    state_extract = True
+    if not paper_text.strip().startswith("Title:\t" + example_data[index_ex]["title"]):
+        print("Mismatch detected.")
+        print(paper_text)
+        return "", "", "", ""
+    example = example_data[index_ex]
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return tasks, gaps, keywords, recent_works
+
+
+# Step 2: Generate Research Hypothesis and Experiment Plan
+def generate_and_store(paper_text, tasks, gaps, keywords, recent_works):
+    if (not state_extract or not state_example or paper_text == ""):
+        return "", "", "", ""
+    global state_generate, index_ex
+    state_generate = True
+    hypothesis = example_data[index_ex]['hypothesis']
+    experiment_plan = example_data[index_ex]['experiment_plan']
+    return hypothesis, experiment_plan, hypothesis, experiment_plan
+
+##########  Phase 2 & 3  ##############
+def start_experiment_agent(hypothesis, plan):
+    if (not state_extract or not state_generate or not state_example):
+        return "", "", ""
+    global state_agent, step_index, state_complete
+    state_agent = True
+    step_index = 0
+    state_complete = False
+    # predefined_message = f"Implement the following hypothesis and experiment plan:\n\nHypothesis:\n{hypothesis}\n\nExperiment Plan:\n{plan}"
+    return example_data[index_ex]['code_init'], predefined_action_log, "", ""
+
+def submit_feedback(user_feedback, history, previous_response):
+    if (not state_extract or not  state_generate or not state_agent or not state_example):
+        return "", "", ""
+    global step_index, state_complete
+    step_index += 1
+    msg = history
+    if step_index < len(process_steps):
+        msg += previous_response + "\nUser feedback:" + user_feedback + "\n\n"
+        response_info = process_steps[step_index]
+        response = info_to_message(response_info)  # Convert dictionary to formatted string
+        response += "Please provide feedback based on the history, response entries, and observation, and questions: "
+        step_index += 1
+        msg += response
+    else:
+        state_complete = True
+        response = "Agent Finished."
+        
+    return msg, response, example_data[index_ex]['code_init'] if state_complete else example_data[index_ex]['code_final'], ""
+
+def load_phase_2_inputs(hypothesis, plan):
+    return hypothesis, plan, "# Code implementation will be displayed here after Start ExperimentAgent."
+
+
+
+predefined_action_log = """
+[Reasoning]: To understand the initial structure and functionality of train.py for effective improvements.
+[Action]: Inspect Script (train.py)
+Input: {"script_name": "train.py", "start_line_number": "1", "end_line_number": "74"}
+Objective: Understand the training script, including data processing, [...]
+[Observation]: The train.py script imports [...]. Sets random seeds [...]. Defines [...] Placeholder functions [...] exist without implementation. [...]
+[Feedback]: The script structure is clear, but key functions (train_model, predict) need proper implementation for proposed model training and prediction.\n
+"""
+
+
+predefined_observation = """
+Epoch [1/10],
+Train MSE: 0.543,
+Test MSE: 0.688
+Epoch [2/10],
+Train MSE: 0.242,
+Test MSE: 0.493\n
+"""
+
+# Initialize the global step_index and history
+process_steps = [
+    {
+        "Action": "Inspect Script Lines (train.py)",
+        "Observation": (
+            "The train.py script imports necessary libraries (e.g., pandas, sklearn, torch). "
+            "Sets random seeds for reproducibility. Defines compute_metrics_for_regression function "
+            "to calculate RMSE for different dimensions. Placeholder functions train_model and "
+            "predict exist without implementations."
+        ),
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The script executed successfully. Generated embeddings using the BERT model. Completed "
+            "the training process without errors. Metrics calculation placeholders indicated areas needing implementation."
+        ),
+    },
+    {
+        "Action": "Edit Script (train.py)",
+        "Observation": (
+            "Edited train.py to separate data loading, model definition, training loop, and evaluation into distinct functions. "
+            "The edited train.py now has clearly defined functions"
+            "for data loading (load_data), model definition (build_model), "
+            "training (train_model), and evaluation (evaluate_model). Similarly, eval.py is reorganized to load the model and perform predictions efficiently."
+        ),
+    },
+    {
+        "Action": "Retrieve Model",
+        "Observation": "CNN and BiLSTM retrieved.",
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The model trained over the specified number of epochs. Training and validation loss values are recorded for each epoch, "
+            "the decrease in loss indicates improved model performance."
+        )
+    },
+    {
+        "Action": "Evaluation",
+        "Observation": predefined_observation,
+    }
+]
+def info_to_message(info):
+    msg = ""
+    for k, v in info.items():
+        if isinstance(v, dict):
+            tempv = v
+            v = ""
+            for k2, v2 in tempv.items():
+                v += f"{k2}:\n  {v2}\n"
+        v = User.indent_text(v, 2)
+        msg += '-' * 64
+        msg += '\n'
+        msg += f"{k}:\n{v}\n"
+    return msg
+
+
+def handle_example_click(example_index):
+    global index_ex
+    index_ex = example_index
+    return load_example(index_ex)  # Simply return the text to display it in the textbox
+
+# Gradio Interface
+with gr.Blocks(css=".gr-examples-label {display: none;}", theme=gr.themes.Default()) as app:
+    gr.Markdown("# MLR- Copilot: Machine Learning Research based on LLM Agents")
+    gr.Markdown("###   ")
+    gr.Markdown("## <span style='color:Orange;'> This UI is for predefined example demo only.</span>")
+    gr.Markdown("## <span style='color:Orange;'> To reproduce the results please use [Github](https://github.com/du-nlp-lab/MLR-Copilot/)</span>")
+
+
+
+    gr.Markdown("MLR-Copilot is a framework where LLMs mimic researchers’ thought processes, designed to enhance the productivity of machine learning research by automating the generation and implementation of research ideas. It begins with a research paper, autonomously generating and validating these ideas, while incorporating human feedback to help reach executable research outcomes.")
+
+
+    gr.Markdown("##    ⬇️  Click an example at bottom to start！")
+
+
+    # Use state variables to store generated hypothesis and experiment plan
+    hypothesis_state = gr.State("")
+    experiment_plan_state = gr.State("")
+
+    ########## Phase 1:  Research Idea Generation Tab ##############
+    with gr.Tab("💡Stage 1: Research Idea Generation"):
+        gr.Markdown("### Extract Research Elements and Generate Research Ideas")
+        with gr.Row():
+            with gr.Column():
+                paper_text_input = gr.Textbox(value="", lines=10, label="📑 Research Paper Text", interactive=False)
+
+                # extract_button = gr.Button("🔍 Extract Research Elements")
+                with gr.Row():
+                    tasks_output = gr.Textbox(placeholder="Research task definition", label="Research Tasks", lines=2, interactive=False)
+                    gaps_output = gr.Textbox(placeholder="Research gaps of current works", label="Research Gaps", lines=2, interactive=False)
+                    keywords_output = gr.Textbox(placeholder="Paper keywords", label="Keywords", lines=2, interactive=False)
+                    recent_works_output = gr.Textbox(placeholder="Recent works extracted from Semantic Scholar", label="Recent Works", lines=2, interactive=False)
+            with gr.Column():
+                with gr.Row():  # Move the button to the top 
+                    generate_button = gr.Button("✍️ Generate Research Hypothesis & Experiment Plan")
+                with gr.Group():
+                    gr.Markdown("### 🌟 Research Idea")
+                    with gr.Row():
+                        hypothesis_output = gr.Textbox(label="Generated Hypothesis", lines=20, interactive=False)
+                        experiment_plan_output = gr.Textbox(label="Generated Experiment Plan", lines=20, interactive=False)
+        
+    
+        # Step 1: Extract Research Elements
+        # extract_button.click(
+        #     fn=extract_research_elements, 
+        #     inputs=paper_text_input, 
+        #     outputs=[tasks_output, gaps_output, keywords_output, recent_works_output]
+        # )
+
+        generate_button.click(
+            fn=generate_and_store, 
+            inputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output], 
+            outputs=[hypothesis_output, experiment_plan_output, hypothesis_state, experiment_plan_state]
+        )
+
+        gr.Examples(
+            examples=example_text,
+            inputs=[paper_text_input],
+            outputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output, hypothesis_output, experiment_plan_output],
+            fn=load_example_and_set_index,
+            run_on_click = True,
+            # label="⬇️ Click an example to load"
+        )
+        
+
+
+
+    ########## Phase 2 & 3:  Experiment implementation and execution ##############
+    with gr.Tab("🧪 Stage 2 & Stage 3: Experiment implementation and execution"):
+        gr.Markdown("### Interact with the ExperimentAgent")
+
+        with gr.Row():
+            with gr.Column():
+                with gr.Group():
+                    gr.Markdown("### 🌟 Generated Research Idea")
+                    with gr.Row():
+                        idea_input = gr.Textbox(label="Generated Research Hypothesis", lines=30, interactive=False)
+                        plan_input = gr.Textbox(label="Generated Experiment Plan", lines=30, interactive=False)
+        
+            with gr.Column():
+                start_exp_agnet = gr.Button("⚙️ Start / Reset ExperimentAgent", elem_classes=["agent-btn"])
+                with gr.Group():
+                    gr.Markdown("### Implementation + Execution Log")    
+                    log = gr.Textbox(label="📖 Execution Log", lines=20, interactive=False)
+                    code_display = gr.Code(label="🧑‍💻 Implementation", language="python", interactive=False)
+
+            with gr.Column():
+                response = gr.Textbox(label="🤖 ExperimentAgent Response", lines=30, interactive=False)
+                feedback = gr.Textbox(placeholder="N/A", label="🧑‍🔬 User Feedback", lines=3, interactive=True)
+                submit_button = gr.Button("Submit", elem_classes=["Submit-btn"])
+        
+        hypothesis_state.change(
+            fn=load_phase_2_inputs, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[idea_input, plan_input, code_display]
+        )
+        
+        # Start research agent
+        start_exp_agnet.click(
+            fn=start_experiment_agent, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[code_display, log, response, feedback]
+        )
+        
+        submit_button.click(
+            fn=submit_feedback,
+            inputs=[feedback, log, response],
+            outputs=[log, response, code_display, feedback]
+        )
+
+# Test
+if __name__ == "__main__":
+    step_index = 0
+    app.launch(share=True)

.history/app_20250404172015.py

@@ -0,0 +1,342 @@
+import gradio as gr
+from pathlib import Path
+from reactagent.environment import Environment
+from reactagent.agents.agent_research import ResearchAgent
+from reactagent.runner import create_parser
+from reactagent import llm
+from reactagent.users.user import User
+import os
+import json
+
+
+# Global variables to store session state
+env = None
+agent = None
+state_example = False
+state_extract = False
+state_generate = False
+state_agent = False
+state_complete = False
+index_ex = "1"
+
+example_text = [
+    "Research Paper 1: Dataset and Baseline for Automatic Student Feedback Analysis",
+    "Research Paper 2: An Empirical Study on the Impact of Code Review on Software Quality"
+]
+
+# Load example JSON file
+def load_example_data():
+    with open("example/example_data.json", "r") as json_file:
+        example_data = json.load(json_file)
+    
+    for idx in example_data.keys():
+        try:
+            file = example_data[idx]["code_init"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_init"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")
+        try:
+            file = example_data[idx]["code_final"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_final"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")     
+    return example_data
+
+example_data = load_example_data()
+
+# Function to handle the selection of an example and populate the respective fields
+def load_example(example_id):
+    global index_ex 
+    index_ex = str(example_id)
+    example = example_data[index_ex]
+    paper_text = 'Title:\t' +  example['title'] + '\n\nAbstract:\t' + example['abstract']
+    global state_extract, index_ex, state_example
+    if not state_example or paper_text == "":
+        return "", "", "", ""
+    state_extract = True
+    if not paper_text.strip().startswith("Title:\t" + example_data[index_ex]["title"]):
+        print("Mismatch detected.")
+        print(paper_text)
+        return "", "", "", ""
+    example = example_data[index_ex]
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return paper_text
+
+example_text = [load_example(1), load_example(2)]
+
+# Function to handle example clicks
+def load_example_and_set_index(paper_text_input):
+    global index_ex, state_example
+    state_example = True
+    index_ex = str(example_text.index(paper_text_input) + 1)
+    paper_text =  load_example(index_ex)
+
+    return paper_text,  "", "", "", "", "", ""
+
+
+
+########## Phase 1 ##############
+
+def extract_research_elements(paper_text):
+    global state_extract, index_ex, state_example
+    if not state_example or paper_text == "":
+        return "", "", "", ""
+    state_extract = True
+    if not paper_text.strip().startswith("Title:\t" + example_data[index_ex]["title"]):
+        print("Mismatch detected.")
+        print(paper_text)
+        return "", "", "", ""
+    example = example_data[index_ex]
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return tasks, gaps, keywords, recent_works
+
+
+# Step 2: Generate Research Hypothesis and Experiment Plan
+def generate_and_store(paper_text, tasks, gaps, keywords, recent_works):
+    if (not state_extract or not state_example or paper_text == ""):
+        return "", "", "", ""
+    global state_generate, index_ex
+    state_generate = True
+    hypothesis = example_data[index_ex]['hypothesis']
+    experiment_plan = example_data[index_ex]['experiment_plan']
+    return hypothesis, experiment_plan, hypothesis, experiment_plan
+
+##########  Phase 2 & 3  ##############
+def start_experiment_agent(hypothesis, plan):
+    if (not state_extract or not state_generate or not state_example):
+        return "", "", ""
+    global state_agent, step_index, state_complete
+    state_agent = True
+    step_index = 0
+    state_complete = False
+    # predefined_message = f"Implement the following hypothesis and experiment plan:\n\nHypothesis:\n{hypothesis}\n\nExperiment Plan:\n{plan}"
+    return example_data[index_ex]['code_init'], predefined_action_log, "", ""
+
+def submit_feedback(user_feedback, history, previous_response):
+    if (not state_extract or not  state_generate or not state_agent or not state_example):
+        return "", "", ""
+    global step_index, state_complete
+    step_index += 1
+    msg = history
+    if step_index < len(process_steps):
+        msg += previous_response + "\nUser feedback:" + user_feedback + "\n\n"
+        response_info = process_steps[step_index]
+        response = info_to_message(response_info)  # Convert dictionary to formatted string
+        response += "Please provide feedback based on the history, response entries, and observation, and questions: "
+        step_index += 1
+        msg += response
+    else:
+        state_complete = True
+        response = "Agent Finished."
+        
+    return msg, response, example_data[index_ex]['code_init'] if state_complete else example_data[index_ex]['code_final'], ""
+
+def load_phase_2_inputs(hypothesis, plan):
+    return hypothesis, plan, "# Code implementation will be displayed here after Start ExperimentAgent."
+
+
+
+predefined_action_log = """
+[Reasoning]: To understand the initial structure and functionality of train.py for effective improvements.
+[Action]: Inspect Script (train.py)
+Input: {"script_name": "train.py", "start_line_number": "1", "end_line_number": "74"}
+Objective: Understand the training script, including data processing, [...]
+[Observation]: The train.py script imports [...]. Sets random seeds [...]. Defines [...] Placeholder functions [...] exist without implementation. [...]
+[Feedback]: The script structure is clear, but key functions (train_model, predict) need proper implementation for proposed model training and prediction.\n
+"""
+
+
+predefined_observation = """
+Epoch [1/10],
+Train MSE: 0.543,
+Test MSE: 0.688
+Epoch [2/10],
+Train MSE: 0.242,
+Test MSE: 0.493\n
+"""
+
+# Initialize the global step_index and history
+process_steps = [
+    {
+        "Action": "Inspect Script Lines (train.py)",
+        "Observation": (
+            "The train.py script imports necessary libraries (e.g., pandas, sklearn, torch). "
+            "Sets random seeds for reproducibility. Defines compute_metrics_for_regression function "
+            "to calculate RMSE for different dimensions. Placeholder functions train_model and "
+            "predict exist without implementations."
+        ),
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The script executed successfully. Generated embeddings using the BERT model. Completed "
+            "the training process without errors. Metrics calculation placeholders indicated areas needing implementation."
+        ),
+    },
+    {
+        "Action": "Edit Script (train.py)",
+        "Observation": (
+            "Edited train.py to separate data loading, model definition, training loop, and evaluation into distinct functions. "
+            "The edited train.py now has clearly defined functions"
+            "for data loading (load_data), model definition (build_model), "
+            "training (train_model), and evaluation (evaluate_model). Similarly, eval.py is reorganized to load the model and perform predictions efficiently."
+        ),
+    },
+    {
+        "Action": "Retrieve Model",
+        "Observation": "CNN and BiLSTM retrieved.",
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The model trained over the specified number of epochs. Training and validation loss values are recorded for each epoch, "
+            "the decrease in loss indicates improved model performance."
+        )
+    },
+    {
+        "Action": "Evaluation",
+        "Observation": predefined_observation,
+    }
+]
+def info_to_message(info):
+    msg = ""
+    for k, v in info.items():
+        if isinstance(v, dict):
+            tempv = v
+            v = ""
+            for k2, v2 in tempv.items():
+                v += f"{k2}:\n  {v2}\n"
+        v = User.indent_text(v, 2)
+        msg += '-' * 64
+        msg += '\n'
+        msg += f"{k}:\n{v}\n"
+    return msg
+
+
+def handle_example_click(example_index):
+    global index_ex
+    index_ex = example_index
+    return load_example(index_ex)  # Simply return the text to display it in the textbox
+
+# Gradio Interface
+with gr.Blocks(css=".gr-examples-label {display: none;}", theme=gr.themes.Default()) as app:
+    gr.Markdown("# MLR- Copilot: Machine Learning Research based on LLM Agents")
+    gr.Markdown("###   ")
+    gr.Markdown("## <span style='color:Orange;'> This UI is for predefined example demo only.</span>")
+    gr.Markdown("## <span style='color:Orange;'> To reproduce the results please use [Github](https://github.com/du-nlp-lab/MLR-Copilot/)</span>")
+
+
+
+    gr.Markdown("MLR-Copilot is a framework where LLMs mimic researchers’ thought processes, designed to enhance the productivity of machine learning research by automating the generation and implementation of research ideas. It begins with a research paper, autonomously generating and validating these ideas, while incorporating human feedback to help reach executable research outcomes.")
+
+
+    gr.Markdown("##    ⬇️  Click an example at bottom to start！")
+
+
+    # Use state variables to store generated hypothesis and experiment plan
+    hypothesis_state = gr.State("")
+    experiment_plan_state = gr.State("")
+
+    ########## Phase 1:  Research Idea Generation Tab ##############
+    with gr.Tab("💡Stage 1: Research Idea Generation"):
+        gr.Markdown("### Extract Research Elements and Generate Research Ideas")
+        with gr.Row():
+            with gr.Column():
+                paper_text_input = gr.Textbox(value="", lines=10, label="📑 Research Paper Text", interactive=False)
+
+                # extract_button = gr.Button("🔍 Extract Research Elements")
+                with gr.Row():
+                    tasks_output = gr.Textbox(placeholder="Research task definition", label="Research Tasks", lines=2, interactive=False)
+                    gaps_output = gr.Textbox(placeholder="Research gaps of current works", label="Research Gaps", lines=2, interactive=False)
+                    keywords_output = gr.Textbox(placeholder="Paper keywords", label="Keywords", lines=2, interactive=False)
+                    recent_works_output = gr.Textbox(placeholder="Recent works extracted from Semantic Scholar", label="Recent Works", lines=2, interactive=False)
+            with gr.Column():
+                with gr.Row():  # Move the button to the top 
+                    generate_button = gr.Button("✍️ Generate Research Hypothesis & Experiment Plan")
+                with gr.Group():
+                    gr.Markdown("### 🌟 Research Idea")
+                    with gr.Row():
+                        hypothesis_output = gr.Textbox(label="Generated Hypothesis", lines=20, interactive=False)
+                        experiment_plan_output = gr.Textbox(label="Generated Experiment Plan", lines=20, interactive=False)
+        
+    
+        # Step 1: Extract Research Elements
+        # extract_button.click(
+        #     fn=extract_research_elements, 
+        #     inputs=paper_text_input, 
+        #     outputs=[tasks_output, gaps_output, keywords_output, recent_works_output]
+        # )
+
+        generate_button.click(
+            fn=generate_and_store, 
+            inputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output], 
+            outputs=[hypothesis_output, experiment_plan_output, hypothesis_state, experiment_plan_state]
+        )
+
+        gr.Examples(
+            examples=example_text,
+            inputs=[paper_text_input],
+            outputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output, hypothesis_output, experiment_plan_output],
+            fn=load_example_and_set_index,
+            run_on_click = True,
+            # label="⬇️ Click an example to load"
+        )
+        
+
+
+
+    ########## Phase 2 & 3:  Experiment implementation and execution ##############
+    with gr.Tab("🧪 Stage 2 & Stage 3: Experiment implementation and execution"):
+        gr.Markdown("### Interact with the ExperimentAgent")
+
+        with gr.Row():
+            with gr.Column():
+                with gr.Group():
+                    gr.Markdown("### 🌟 Generated Research Idea")
+                    with gr.Row():
+                        idea_input = gr.Textbox(label="Generated Research Hypothesis", lines=30, interactive=False)
+                        plan_input = gr.Textbox(label="Generated Experiment Plan", lines=30, interactive=False)
+        
+            with gr.Column():
+                start_exp_agnet = gr.Button("⚙️ Start / Reset ExperimentAgent", elem_classes=["agent-btn"])
+                with gr.Group():
+                    gr.Markdown("### Implementation + Execution Log")    
+                    log = gr.Textbox(label="📖 Execution Log", lines=20, interactive=False)
+                    code_display = gr.Code(label="🧑‍💻 Implementation", language="python", interactive=False)
+
+            with gr.Column():
+                response = gr.Textbox(label="🤖 ExperimentAgent Response", lines=30, interactive=False)
+                feedback = gr.Textbox(placeholder="N/A", label="🧑‍🔬 User Feedback", lines=3, interactive=True)
+                submit_button = gr.Button("Submit", elem_classes=["Submit-btn"])
+        
+        hypothesis_state.change(
+            fn=load_phase_2_inputs, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[idea_input, plan_input, code_display]
+        )
+        
+        # Start research agent
+        start_exp_agnet.click(
+            fn=start_experiment_agent, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[code_display, log, response, feedback]
+        )
+        
+        submit_button.click(
+            fn=submit_feedback,
+            inputs=[feedback, log, response],
+            outputs=[log, response, code_display, feedback]
+        )
+
+# Test
+if __name__ == "__main__":
+    step_index = 0
+    app.launch(share=True)

.history/app_20250404172040.py

@@ -0,0 +1,340 @@
+import gradio as gr
+from pathlib import Path
+from reactagent.environment import Environment
+from reactagent.agents.agent_research import ResearchAgent
+from reactagent.runner import create_parser
+from reactagent import llm
+from reactagent.users.user import User
+import os
+import json
+
+
+# Global variables to store session state
+env = None
+agent = None
+state_example = False
+state_extract = False
+state_generate = False
+state_agent = False
+state_complete = False
+index_ex = "1"
+
+example_text = [
+    "Research Paper 1: Dataset and Baseline for Automatic Student Feedback Analysis",
+    "Research Paper 2: An Empirical Study on the Impact of Code Review on Software Quality"
+]
+
+# Load example JSON file
+def load_example_data():
+    with open("example/example_data.json", "r") as json_file:
+        example_data = json.load(json_file)
+    
+    for idx in example_data.keys():
+        try:
+            file = example_data[idx]["code_init"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_init"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")
+        try:
+            file = example_data[idx]["code_final"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_final"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")     
+    return example_data
+
+example_data = load_example_data()
+
+# Function to handle the selection of an example and populate the respective fields
+def load_example(example_id):
+    global index_ex 
+    index_ex = str(example_id)
+    example = example_data[index_ex]
+    paper_text = 'Title:\t' +  example['title'] + '\n\nAbstract:\t' + example['abstract']
+    global state_extract, index_ex, state_example
+    state_extract = True
+    if not paper_text.strip().startswith("Title:\t" + example_data[index_ex]["title"]):
+        print("Mismatch detected.")
+        print(paper_text)
+        return "", "", "", ""
+    example = example_data[index_ex]
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return paper_text
+
+example_text = [load_example(1), load_example(2)]
+
+# Function to handle example clicks
+def load_example_and_set_index(paper_text_input):
+    global index_ex, state_example
+    state_example = True
+    index_ex = str(example_text.index(paper_text_input) + 1)
+    paper_text =  load_example(index_ex)
+
+    return paper_text,  "", "", "", "", "", ""
+
+
+
+########## Phase 1 ##############
+
+def extract_research_elements(paper_text):
+    global state_extract, index_ex, state_example
+    if not state_example or paper_text == "":
+        return "", "", "", ""
+    state_extract = True
+    if not paper_text.strip().startswith("Title:\t" + example_data[index_ex]["title"]):
+        print("Mismatch detected.")
+        print(paper_text)
+        return "", "", "", ""
+    example = example_data[index_ex]
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return tasks, gaps, keywords, recent_works
+
+
+# Step 2: Generate Research Hypothesis and Experiment Plan
+def generate_and_store(paper_text, tasks, gaps, keywords, recent_works):
+    if (not state_extract or not state_example or paper_text == ""):
+        return "", "", "", ""
+    global state_generate, index_ex
+    state_generate = True
+    hypothesis = example_data[index_ex]['hypothesis']
+    experiment_plan = example_data[index_ex]['experiment_plan']
+    return hypothesis, experiment_plan, hypothesis, experiment_plan
+
+##########  Phase 2 & 3  ##############
+def start_experiment_agent(hypothesis, plan):
+    if (not state_extract or not state_generate or not state_example):
+        return "", "", ""
+    global state_agent, step_index, state_complete
+    state_agent = True
+    step_index = 0
+    state_complete = False
+    # predefined_message = f"Implement the following hypothesis and experiment plan:\n\nHypothesis:\n{hypothesis}\n\nExperiment Plan:\n{plan}"
+    return example_data[index_ex]['code_init'], predefined_action_log, "", ""
+
+def submit_feedback(user_feedback, history, previous_response):
+    if (not state_extract or not  state_generate or not state_agent or not state_example):
+        return "", "", ""
+    global step_index, state_complete
+    step_index += 1
+    msg = history
+    if step_index < len(process_steps):
+        msg += previous_response + "\nUser feedback:" + user_feedback + "\n\n"
+        response_info = process_steps[step_index]
+        response = info_to_message(response_info)  # Convert dictionary to formatted string
+        response += "Please provide feedback based on the history, response entries, and observation, and questions: "
+        step_index += 1
+        msg += response
+    else:
+        state_complete = True
+        response = "Agent Finished."
+        
+    return msg, response, example_data[index_ex]['code_init'] if state_complete else example_data[index_ex]['code_final'], ""
+
+def load_phase_2_inputs(hypothesis, plan):
+    return hypothesis, plan, "# Code implementation will be displayed here after Start ExperimentAgent."
+
+
+
+predefined_action_log = """
+[Reasoning]: To understand the initial structure and functionality of train.py for effective improvements.
+[Action]: Inspect Script (train.py)
+Input: {"script_name": "train.py", "start_line_number": "1", "end_line_number": "74"}
+Objective: Understand the training script, including data processing, [...]
+[Observation]: The train.py script imports [...]. Sets random seeds [...]. Defines [...] Placeholder functions [...] exist without implementation. [...]
+[Feedback]: The script structure is clear, but key functions (train_model, predict) need proper implementation for proposed model training and prediction.\n
+"""
+
+
+predefined_observation = """
+Epoch [1/10],
+Train MSE: 0.543,
+Test MSE: 0.688
+Epoch [2/10],
+Train MSE: 0.242,
+Test MSE: 0.493\n
+"""
+
+# Initialize the global step_index and history
+process_steps = [
+    {
+        "Action": "Inspect Script Lines (train.py)",
+        "Observation": (
+            "The train.py script imports necessary libraries (e.g., pandas, sklearn, torch). "
+            "Sets random seeds for reproducibility. Defines compute_metrics_for_regression function "
+            "to calculate RMSE for different dimensions. Placeholder functions train_model and "
+            "predict exist without implementations."
+        ),
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The script executed successfully. Generated embeddings using the BERT model. Completed "
+            "the training process without errors. Metrics calculation placeholders indicated areas needing implementation."
+        ),
+    },
+    {
+        "Action": "Edit Script (train.py)",
+        "Observation": (
+            "Edited train.py to separate data loading, model definition, training loop, and evaluation into distinct functions. "
+            "The edited train.py now has clearly defined functions"
+            "for data loading (load_data), model definition (build_model), "
+            "training (train_model), and evaluation (evaluate_model). Similarly, eval.py is reorganized to load the model and perform predictions efficiently."
+        ),
+    },
+    {
+        "Action": "Retrieve Model",
+        "Observation": "CNN and BiLSTM retrieved.",
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The model trained over the specified number of epochs. Training and validation loss values are recorded for each epoch, "
+            "the decrease in loss indicates improved model performance."
+        )
+    },
+    {
+        "Action": "Evaluation",
+        "Observation": predefined_observation,
+    }
+]
+def info_to_message(info):
+    msg = ""
+    for k, v in info.items():
+        if isinstance(v, dict):
+            tempv = v
+            v = ""
+            for k2, v2 in tempv.items():
+                v += f"{k2}:\n  {v2}\n"
+        v = User.indent_text(v, 2)
+        msg += '-' * 64
+        msg += '\n'
+        msg += f"{k}:\n{v}\n"
+    return msg
+
+
+def handle_example_click(example_index):
+    global index_ex
+    index_ex = example_index
+    return load_example(index_ex)  # Simply return the text to display it in the textbox
+
+# Gradio Interface
+with gr.Blocks(css=".gr-examples-label {display: none;}", theme=gr.themes.Default()) as app:
+    gr.Markdown("# MLR- Copilot: Machine Learning Research based on LLM Agents")
+    gr.Markdown("###   ")
+    gr.Markdown("## <span style='color:Orange;'> This UI is for predefined example demo only.</span>")
+    gr.Markdown("## <span style='color:Orange;'> To reproduce the results please use [Github](https://github.com/du-nlp-lab/MLR-Copilot/)</span>")
+
+
+
+    gr.Markdown("MLR-Copilot is a framework where LLMs mimic researchers’ thought processes, designed to enhance the productivity of machine learning research by automating the generation and implementation of research ideas. It begins with a research paper, autonomously generating and validating these ideas, while incorporating human feedback to help reach executable research outcomes.")
+
+
+    gr.Markdown("##    ⬇️  Click an example at bottom to start！")
+
+
+    # Use state variables to store generated hypothesis and experiment plan
+    hypothesis_state = gr.State("")
+    experiment_plan_state = gr.State("")
+
+    ########## Phase 1:  Research Idea Generation Tab ##############
+    with gr.Tab("💡Stage 1: Research Idea Generation"):
+        gr.Markdown("### Extract Research Elements and Generate Research Ideas")
+        with gr.Row():
+            with gr.Column():
+                paper_text_input = gr.Textbox(value="", lines=10, label="📑 Research Paper Text", interactive=False)
+
+                # extract_button = gr.Button("🔍 Extract Research Elements")
+                with gr.Row():
+                    tasks_output = gr.Textbox(placeholder="Research task definition", label="Research Tasks", lines=2, interactive=False)
+                    gaps_output = gr.Textbox(placeholder="Research gaps of current works", label="Research Gaps", lines=2, interactive=False)
+                    keywords_output = gr.Textbox(placeholder="Paper keywords", label="Keywords", lines=2, interactive=False)
+                    recent_works_output = gr.Textbox(placeholder="Recent works extracted from Semantic Scholar", label="Recent Works", lines=2, interactive=False)
+            with gr.Column():
+                with gr.Row():  # Move the button to the top 
+                    generate_button = gr.Button("✍️ Generate Research Hypothesis & Experiment Plan")
+                with gr.Group():
+                    gr.Markdown("### 🌟 Research Idea")
+                    with gr.Row():
+                        hypothesis_output = gr.Textbox(label="Generated Hypothesis", lines=20, interactive=False)
+                        experiment_plan_output = gr.Textbox(label="Generated Experiment Plan", lines=20, interactive=False)
+        
+    
+        # Step 1: Extract Research Elements
+        # extract_button.click(
+        #     fn=extract_research_elements, 
+        #     inputs=paper_text_input, 
+        #     outputs=[tasks_output, gaps_output, keywords_output, recent_works_output]
+        # )
+
+        generate_button.click(
+            fn=generate_and_store, 
+            inputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output], 
+            outputs=[hypothesis_output, experiment_plan_output, hypothesis_state, experiment_plan_state]
+        )
+
+        gr.Examples(
+            examples=example_text,
+            inputs=[paper_text_input],
+            outputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output, hypothesis_output, experiment_plan_output],
+            fn=load_example_and_set_index,
+            run_on_click = True,
+            # label="⬇️ Click an example to load"
+        )
+        
+
+
+
+    ########## Phase 2 & 3:  Experiment implementation and execution ##############
+    with gr.Tab("🧪 Stage 2 & Stage 3: Experiment implementation and execution"):
+        gr.Markdown("### Interact with the ExperimentAgent")
+
+        with gr.Row():
+            with gr.Column():
+                with gr.Group():
+                    gr.Markdown("### 🌟 Generated Research Idea")
+                    with gr.Row():
+                        idea_input = gr.Textbox(label="Generated Research Hypothesis", lines=30, interactive=False)
+                        plan_input = gr.Textbox(label="Generated Experiment Plan", lines=30, interactive=False)
+        
+            with gr.Column():
+                start_exp_agnet = gr.Button("⚙️ Start / Reset ExperimentAgent", elem_classes=["agent-btn"])
+                with gr.Group():
+                    gr.Markdown("### Implementation + Execution Log")    
+                    log = gr.Textbox(label="📖 Execution Log", lines=20, interactive=False)
+                    code_display = gr.Code(label="🧑‍💻 Implementation", language="python", interactive=False)
+
+            with gr.Column():
+                response = gr.Textbox(label="🤖 ExperimentAgent Response", lines=30, interactive=False)
+                feedback = gr.Textbox(placeholder="N/A", label="🧑‍🔬 User Feedback", lines=3, interactive=True)
+                submit_button = gr.Button("Submit", elem_classes=["Submit-btn"])
+        
+        hypothesis_state.change(
+            fn=load_phase_2_inputs, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[idea_input, plan_input, code_display]
+        )
+        
+        # Start research agent
+        start_exp_agnet.click(
+            fn=start_experiment_agent, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[code_display, log, response, feedback]
+        )
+        
+        submit_button.click(
+            fn=submit_feedback,
+            inputs=[feedback, log, response],
+            outputs=[log, response, code_display, feedback]
+        )
+
+# Test
+if __name__ == "__main__":
+    step_index = 0
+    app.launch(share=True)

.history/app_20250404172125.py

@@ -0,0 +1,333 @@
+import gradio as gr
+from pathlib import Path
+from reactagent.environment import Environment
+from reactagent.agents.agent_research import ResearchAgent
+from reactagent.runner import create_parser
+from reactagent import llm
+from reactagent.users.user import User
+import os
+import json
+
+
+# Global variables to store session state
+env = None
+agent = None
+state_example = False
+state_extract = False
+state_generate = False
+state_agent = False
+state_complete = False
+index_ex = "1"
+
+example_text = [
+    "Research Paper 1: Dataset and Baseline for Automatic Student Feedback Analysis",
+    "Research Paper 2: An Empirical Study on the Impact of Code Review on Software Quality"
+]
+
+# Load example JSON file
+def load_example_data():
+    with open("example/example_data.json", "r") as json_file:
+        example_data = json.load(json_file)
+    
+    for idx in example_data.keys():
+        try:
+            file = example_data[idx]["code_init"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_init"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")
+        try:
+            file = example_data[idx]["code_final"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_final"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")     
+    return example_data
+
+example_data = load_example_data()
+
+# Function to handle the selection of an example and populate the respective fields
+def load_example(example_id):
+    global index_ex 
+    index_ex = str(example_id)
+    example = example_data[index_ex]
+    paper_text = 'Title:\t' +  example['title'] + '\n\nAbstract:\t' + example['abstract']
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return paper_text, tasks, gaps, keywords, recent_works
+
+example_text = [load_example(1), load_example(2)]
+
+# Function to handle example clicks
+def load_example_and_set_index(paper_text_input):
+    global index_ex, state_example
+    state_example = True
+    index_ex = str(example_text.index(paper_text_input) + 1)
+    paper_text =  load_example(index_ex)
+
+    return paper_text,  "", "", "", "", "", ""
+
+
+
+########## Phase 1 ##############
+
+def extract_research_elements(paper_text):
+    global state_extract, index_ex, state_example
+    if not state_example or paper_text == "":
+        return "", "", "", ""
+    state_extract = True
+    if not paper_text.strip().startswith("Title:\t" + example_data[index_ex]["title"]):
+        print("Mismatch detected.")
+        print(paper_text)
+        return "", "", "", ""
+    example = example_data[index_ex]
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return tasks, gaps, keywords, recent_works
+
+
+# Step 2: Generate Research Hypothesis and Experiment Plan
+def generate_and_store(paper_text, tasks, gaps, keywords, recent_works):
+    if (not state_extract or not state_example or paper_text == ""):
+        return "", "", "", ""
+    global state_generate, index_ex
+    state_generate = True
+    hypothesis = example_data[index_ex]['hypothesis']
+    experiment_plan = example_data[index_ex]['experiment_plan']
+    return hypothesis, experiment_plan, hypothesis, experiment_plan
+
+##########  Phase 2 & 3  ##############
+def start_experiment_agent(hypothesis, plan):
+    if (not state_extract or not state_generate or not state_example):
+        return "", "", ""
+    global state_agent, step_index, state_complete
+    state_agent = True
+    step_index = 0
+    state_complete = False
+    # predefined_message = f"Implement the following hypothesis and experiment plan:\n\nHypothesis:\n{hypothesis}\n\nExperiment Plan:\n{plan}"
+    return example_data[index_ex]['code_init'], predefined_action_log, "", ""
+
+def submit_feedback(user_feedback, history, previous_response):
+    if (not state_extract or not  state_generate or not state_agent or not state_example):
+        return "", "", ""
+    global step_index, state_complete
+    step_index += 1
+    msg = history
+    if step_index < len(process_steps):
+        msg += previous_response + "\nUser feedback:" + user_feedback + "\n\n"
+        response_info = process_steps[step_index]
+        response = info_to_message(response_info)  # Convert dictionary to formatted string
+        response += "Please provide feedback based on the history, response entries, and observation, and questions: "
+        step_index += 1
+        msg += response
+    else:
+        state_complete = True
+        response = "Agent Finished."
+        
+    return msg, response, example_data[index_ex]['code_init'] if state_complete else example_data[index_ex]['code_final'], ""
+
+def load_phase_2_inputs(hypothesis, plan):
+    return hypothesis, plan, "# Code implementation will be displayed here after Start ExperimentAgent."
+
+
+
+predefined_action_log = """
+[Reasoning]: To understand the initial structure and functionality of train.py for effective improvements.
+[Action]: Inspect Script (train.py)
+Input: {"script_name": "train.py", "start_line_number": "1", "end_line_number": "74"}
+Objective: Understand the training script, including data processing, [...]
+[Observation]: The train.py script imports [...]. Sets random seeds [...]. Defines [...] Placeholder functions [...] exist without implementation. [...]
+[Feedback]: The script structure is clear, but key functions (train_model, predict) need proper implementation for proposed model training and prediction.\n
+"""
+
+
+predefined_observation = """
+Epoch [1/10],
+Train MSE: 0.543,
+Test MSE: 0.688
+Epoch [2/10],
+Train MSE: 0.242,
+Test MSE: 0.493\n
+"""
+
+# Initialize the global step_index and history
+process_steps = [
+    {
+        "Action": "Inspect Script Lines (train.py)",
+        "Observation": (
+            "The train.py script imports necessary libraries (e.g., pandas, sklearn, torch). "
+            "Sets random seeds for reproducibility. Defines compute_metrics_for_regression function "
+            "to calculate RMSE for different dimensions. Placeholder functions train_model and "
+            "predict exist without implementations."
+        ),
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The script executed successfully. Generated embeddings using the BERT model. Completed "
+            "the training process without errors. Metrics calculation placeholders indicated areas needing implementation."
+        ),
+    },
+    {
+        "Action": "Edit Script (train.py)",
+        "Observation": (
+            "Edited train.py to separate data loading, model definition, training loop, and evaluation into distinct functions. "
+            "The edited train.py now has clearly defined functions"
+            "for data loading (load_data), model definition (build_model), "
+            "training (train_model), and evaluation (evaluate_model). Similarly, eval.py is reorganized to load the model and perform predictions efficiently."
+        ),
+    },
+    {
+        "Action": "Retrieve Model",
+        "Observation": "CNN and BiLSTM retrieved.",
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The model trained over the specified number of epochs. Training and validation loss values are recorded for each epoch, "
+            "the decrease in loss indicates improved model performance."
+        )
+    },
+    {
+        "Action": "Evaluation",
+        "Observation": predefined_observation,
+    }
+]
+def info_to_message(info):
+    msg = ""
+    for k, v in info.items():
+        if isinstance(v, dict):
+            tempv = v
+            v = ""
+            for k2, v2 in tempv.items():
+                v += f"{k2}:\n  {v2}\n"
+        v = User.indent_text(v, 2)
+        msg += '-' * 64
+        msg += '\n'
+        msg += f"{k}:\n{v}\n"
+    return msg
+
+
+def handle_example_click(example_index):
+    global index_ex
+    index_ex = example_index
+    return load_example(index_ex)  # Simply return the text to display it in the textbox
+
+# Gradio Interface
+with gr.Blocks(css=".gr-examples-label {display: none;}", theme=gr.themes.Default()) as app:
+    gr.Markdown("# MLR- Copilot: Machine Learning Research based on LLM Agents")
+    gr.Markdown("###   ")
+    gr.Markdown("## <span style='color:Orange;'> This UI is for predefined example demo only.</span>")
+    gr.Markdown("## <span style='color:Orange;'> To reproduce the results please use [Github](https://github.com/du-nlp-lab/MLR-Copilot/)</span>")
+
+
+
+    gr.Markdown("MLR-Copilot is a framework where LLMs mimic researchers’ thought processes, designed to enhance the productivity of machine learning research by automating the generation and implementation of research ideas. It begins with a research paper, autonomously generating and validating these ideas, while incorporating human feedback to help reach executable research outcomes.")
+
+
+    gr.Markdown("##    ⬇️  Click an example at bottom to start！")
+
+
+    # Use state variables to store generated hypothesis and experiment plan
+    hypothesis_state = gr.State("")
+    experiment_plan_state = gr.State("")
+
+    ########## Phase 1:  Research Idea Generation Tab ##############
+    with gr.Tab("💡Stage 1: Research Idea Generation"):
+        gr.Markdown("### Extract Research Elements and Generate Research Ideas")
+        with gr.Row():
+            with gr.Column():
+                paper_text_input = gr.Textbox(value="", lines=10, label="📑 Research Paper Text", interactive=False)
+
+                # extract_button = gr.Button("🔍 Extract Research Elements")
+                with gr.Row():
+                    tasks_output = gr.Textbox(placeholder="Research task definition", label="Research Tasks", lines=2, interactive=False)
+                    gaps_output = gr.Textbox(placeholder="Research gaps of current works", label="Research Gaps", lines=2, interactive=False)
+                    keywords_output = gr.Textbox(placeholder="Paper keywords", label="Keywords", lines=2, interactive=False)
+                    recent_works_output = gr.Textbox(placeholder="Recent works extracted from Semantic Scholar", label="Recent Works", lines=2, interactive=False)
+            with gr.Column():
+                with gr.Row():  # Move the button to the top 
+                    generate_button = gr.Button("✍️ Generate Research Hypothesis & Experiment Plan")
+                with gr.Group():
+                    gr.Markdown("### 🌟 Research Idea")
+                    with gr.Row():
+                        hypothesis_output = gr.Textbox(label="Generated Hypothesis", lines=20, interactive=False)
+                        experiment_plan_output = gr.Textbox(label="Generated Experiment Plan", lines=20, interactive=False)
+        
+    
+        # Step 1: Extract Research Elements
+        # extract_button.click(
+        #     fn=extract_research_elements, 
+        #     inputs=paper_text_input, 
+        #     outputs=[tasks_output, gaps_output, keywords_output, recent_works_output]
+        # )
+
+        generate_button.click(
+            fn=generate_and_store, 
+            inputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output], 
+            outputs=[hypothesis_output, experiment_plan_output, hypothesis_state, experiment_plan_state]
+        )
+
+        gr.Examples(
+            examples=example_text,
+            inputs=[paper_text_input],
+            outputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output, hypothesis_output, experiment_plan_output],
+            fn=load_example_and_set_index,
+            run_on_click = True,
+            # label="⬇️ Click an example to load"
+        )
+        
+
+
+
+    ########## Phase 2 & 3:  Experiment implementation and execution ##############
+    with gr.Tab("🧪 Stage 2 & Stage 3: Experiment implementation and execution"):
+        gr.Markdown("### Interact with the ExperimentAgent")
+
+        with gr.Row():
+            with gr.Column():
+                with gr.Group():
+                    gr.Markdown("### 🌟 Generated Research Idea")
+                    with gr.Row():
+                        idea_input = gr.Textbox(label="Generated Research Hypothesis", lines=30, interactive=False)
+                        plan_input = gr.Textbox(label="Generated Experiment Plan", lines=30, interactive=False)
+        
+            with gr.Column():
+                start_exp_agnet = gr.Button("⚙️ Start / Reset ExperimentAgent", elem_classes=["agent-btn"])
+                with gr.Group():
+                    gr.Markdown("### Implementation + Execution Log")    
+                    log = gr.Textbox(label="📖 Execution Log", lines=20, interactive=False)
+                    code_display = gr.Code(label="🧑‍💻 Implementation", language="python", interactive=False)
+
+            with gr.Column():
+                response = gr.Textbox(label="🤖 ExperimentAgent Response", lines=30, interactive=False)
+                feedback = gr.Textbox(placeholder="N/A", label="🧑‍🔬 User Feedback", lines=3, interactive=True)
+                submit_button = gr.Button("Submit", elem_classes=["Submit-btn"])
+        
+        hypothesis_state.change(
+            fn=load_phase_2_inputs, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[idea_input, plan_input, code_display]
+        )
+        
+        # Start research agent
+        start_exp_agnet.click(
+            fn=start_experiment_agent, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[code_display, log, response, feedback]
+        )
+        
+        submit_button.click(
+            fn=submit_feedback,
+            inputs=[feedback, log, response],
+            outputs=[log, response, code_display, feedback]
+        )
+
+# Test
+if __name__ == "__main__":
+    step_index = 0
+    app.launch(share=True)

.history/app_20250404172143.py

@@ -0,0 +1,333 @@
+import gradio as gr
+from pathlib import Path
+from reactagent.environment import Environment
+from reactagent.agents.agent_research import ResearchAgent
+from reactagent.runner import create_parser
+from reactagent import llm
+from reactagent.users.user import User
+import os
+import json
+
+
+# Global variables to store session state
+env = None
+agent = None
+state_example = False
+state_extract = False
+state_generate = False
+state_agent = False
+state_complete = False
+index_ex = "1"
+
+example_text = [
+    "Research Paper 1: Dataset and Baseline for Automatic Student Feedback Analysis",
+    "Research Paper 2: An Empirical Study on the Impact of Code Review on Software Quality"
+]
+
+# Load example JSON file
+def load_example_data():
+    with open("example/example_data.json", "r") as json_file:
+        example_data = json.load(json_file)
+    
+    for idx in example_data.keys():
+        try:
+            file = example_data[idx]["code_init"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_init"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")
+        try:
+            file = example_data[idx]["code_final"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_final"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")     
+    return example_data
+
+example_data = load_example_data()
+
+# Function to handle the selection of an example and populate the respective fields
+def load_example(example_id):
+    global index_ex 
+    index_ex = str(example_id)
+    example = example_data[index_ex]
+    paper_text = 'Title:\t' +  example['title'] + '\n\nAbstract:\t' + example['abstract']
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return [paper_text, tasks, gaps, keywords, recent_works]
+
+example_text = [load_example(1), load_example(2)]
+
+# Function to handle example clicks
+def load_example_and_set_index(paper_text_input):
+    global index_ex, state_example
+    state_example = True
+    index_ex = str(example_text.index(paper_text_input) + 1)
+    paper_text =  load_example(index_ex)
+
+    return paper_text,  "", "", "", "", "", ""
+
+
+
+########## Phase 1 ##############
+
+def extract_research_elements(paper_text):
+    global state_extract, index_ex, state_example
+    if not state_example or paper_text == "":
+        return "", "", "", ""
+    state_extract = True
+    if not paper_text.strip().startswith("Title:\t" + example_data[index_ex]["title"]):
+        print("Mismatch detected.")
+        print(paper_text)
+        return "", "", "", ""
+    example = example_data[index_ex]
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return tasks, gaps, keywords, recent_works
+
+
+# Step 2: Generate Research Hypothesis and Experiment Plan
+def generate_and_store(paper_text, tasks, gaps, keywords, recent_works):
+    if (not state_extract or not state_example or paper_text == ""):
+        return "", "", "", ""
+    global state_generate, index_ex
+    state_generate = True
+    hypothesis = example_data[index_ex]['hypothesis']
+    experiment_plan = example_data[index_ex]['experiment_plan']
+    return hypothesis, experiment_plan, hypothesis, experiment_plan
+
+##########  Phase 2 & 3  ##############
+def start_experiment_agent(hypothesis, plan):
+    if (not state_extract or not state_generate or not state_example):
+        return "", "", ""
+    global state_agent, step_index, state_complete
+    state_agent = True
+    step_index = 0
+    state_complete = False
+    # predefined_message = f"Implement the following hypothesis and experiment plan:\n\nHypothesis:\n{hypothesis}\n\nExperiment Plan:\n{plan}"
+    return example_data[index_ex]['code_init'], predefined_action_log, "", ""
+
+def submit_feedback(user_feedback, history, previous_response):
+    if (not state_extract or not  state_generate or not state_agent or not state_example):
+        return "", "", ""
+    global step_index, state_complete
+    step_index += 1
+    msg = history
+    if step_index < len(process_steps):
+        msg += previous_response + "\nUser feedback:" + user_feedback + "\n\n"
+        response_info = process_steps[step_index]
+        response = info_to_message(response_info)  # Convert dictionary to formatted string
+        response += "Please provide feedback based on the history, response entries, and observation, and questions: "
+        step_index += 1
+        msg += response
+    else:
+        state_complete = True
+        response = "Agent Finished."
+        
+    return msg, response, example_data[index_ex]['code_init'] if state_complete else example_data[index_ex]['code_final'], ""
+
+def load_phase_2_inputs(hypothesis, plan):
+    return hypothesis, plan, "# Code implementation will be displayed here after Start ExperimentAgent."
+
+
+
+predefined_action_log = """
+[Reasoning]: To understand the initial structure and functionality of train.py for effective improvements.
+[Action]: Inspect Script (train.py)
+Input: {"script_name": "train.py", "start_line_number": "1", "end_line_number": "74"}
+Objective: Understand the training script, including data processing, [...]
+[Observation]: The train.py script imports [...]. Sets random seeds [...]. Defines [...] Placeholder functions [...] exist without implementation. [...]
+[Feedback]: The script structure is clear, but key functions (train_model, predict) need proper implementation for proposed model training and prediction.\n
+"""
+
+
+predefined_observation = """
+Epoch [1/10],
+Train MSE: 0.543,
+Test MSE: 0.688
+Epoch [2/10],
+Train MSE: 0.242,
+Test MSE: 0.493\n
+"""
+
+# Initialize the global step_index and history
+process_steps = [
+    {
+        "Action": "Inspect Script Lines (train.py)",
+        "Observation": (
+            "The train.py script imports necessary libraries (e.g., pandas, sklearn, torch). "
+            "Sets random seeds for reproducibility. Defines compute_metrics_for_regression function "
+            "to calculate RMSE for different dimensions. Placeholder functions train_model and "
+            "predict exist without implementations."
+        ),
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The script executed successfully. Generated embeddings using the BERT model. Completed "
+            "the training process without errors. Metrics calculation placeholders indicated areas needing implementation."
+        ),
+    },
+    {
+        "Action": "Edit Script (train.py)",
+        "Observation": (
+            "Edited train.py to separate data loading, model definition, training loop, and evaluation into distinct functions. "
+            "The edited train.py now has clearly defined functions"
+            "for data loading (load_data), model definition (build_model), "
+            "training (train_model), and evaluation (evaluate_model). Similarly, eval.py is reorganized to load the model and perform predictions efficiently."
+        ),
+    },
+    {
+        "Action": "Retrieve Model",
+        "Observation": "CNN and BiLSTM retrieved.",
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The model trained over the specified number of epochs. Training and validation loss values are recorded for each epoch, "
+            "the decrease in loss indicates improved model performance."
+        )
+    },
+    {
+        "Action": "Evaluation",
+        "Observation": predefined_observation,
+    }
+]
+def info_to_message(info):
+    msg = ""
+    for k, v in info.items():
+        if isinstance(v, dict):
+            tempv = v
+            v = ""
+            for k2, v2 in tempv.items():
+                v += f"{k2}:\n  {v2}\n"
+        v = User.indent_text(v, 2)
+        msg += '-' * 64
+        msg += '\n'
+        msg += f"{k}:\n{v}\n"
+    return msg
+
+
+def handle_example_click(example_index):
+    global index_ex
+    index_ex = example_index
+    return load_example(index_ex)  # Simply return the text to display it in the textbox
+
+# Gradio Interface
+with gr.Blocks(css=".gr-examples-label {display: none;}", theme=gr.themes.Default()) as app:
+    gr.Markdown("# MLR- Copilot: Machine Learning Research based on LLM Agents")
+    gr.Markdown("###   ")
+    gr.Markdown("## <span style='color:Orange;'> This UI is for predefined example demo only.</span>")
+    gr.Markdown("## <span style='color:Orange;'> To reproduce the results please use [Github](https://github.com/du-nlp-lab/MLR-Copilot/)</span>")
+
+
+
+    gr.Markdown("MLR-Copilot is a framework where LLMs mimic researchers’ thought processes, designed to enhance the productivity of machine learning research by automating the generation and implementation of research ideas. It begins with a research paper, autonomously generating and validating these ideas, while incorporating human feedback to help reach executable research outcomes.")
+
+
+    gr.Markdown("##    ⬇️  Click an example at bottom to start！")
+
+
+    # Use state variables to store generated hypothesis and experiment plan
+    hypothesis_state = gr.State("")
+    experiment_plan_state = gr.State("")
+
+    ########## Phase 1:  Research Idea Generation Tab ##############
+    with gr.Tab("💡Stage 1: Research Idea Generation"):
+        gr.Markdown("### Extract Research Elements and Generate Research Ideas")
+        with gr.Row():
+            with gr.Column():
+                paper_text_input = gr.Textbox(value="", lines=10, label="📑 Research Paper Text", interactive=False)
+
+                # extract_button = gr.Button("🔍 Extract Research Elements")
+                with gr.Row():
+                    tasks_output = gr.Textbox(placeholder="Research task definition", label="Research Tasks", lines=2, interactive=False)
+                    gaps_output = gr.Textbox(placeholder="Research gaps of current works", label="Research Gaps", lines=2, interactive=False)
+                    keywords_output = gr.Textbox(placeholder="Paper keywords", label="Keywords", lines=2, interactive=False)
+                    recent_works_output = gr.Textbox(placeholder="Recent works extracted from Semantic Scholar", label="Recent Works", lines=2, interactive=False)
+            with gr.Column():
+                with gr.Row():  # Move the button to the top 
+                    generate_button = gr.Button("✍️ Generate Research Hypothesis & Experiment Plan")
+                with gr.Group():
+                    gr.Markdown("### 🌟 Research Idea")
+                    with gr.Row():
+                        hypothesis_output = gr.Textbox(label="Generated Hypothesis", lines=20, interactive=False)
+                        experiment_plan_output = gr.Textbox(label="Generated Experiment Plan", lines=20, interactive=False)
+        
+    
+        # Step 1: Extract Research Elements
+        # extract_button.click(
+        #     fn=extract_research_elements, 
+        #     inputs=paper_text_input, 
+        #     outputs=[tasks_output, gaps_output, keywords_output, recent_works_output]
+        # )
+
+        generate_button.click(
+            fn=generate_and_store, 
+            inputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output], 
+            outputs=[hypothesis_output, experiment_plan_output, hypothesis_state, experiment_plan_state]
+        )
+
+        gr.Examples(
+            examples=example_text,
+            inputs=[paper_text_input],
+            outputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output, hypothesis_output, experiment_plan_output],
+            fn=load_example_and_set_index,
+            run_on_click = True,
+            # label="⬇️ Click an example to load"
+        )
+        
+
+
+
+    ########## Phase 2 & 3:  Experiment implementation and execution ##############
+    with gr.Tab("🧪 Stage 2 & Stage 3: Experiment implementation and execution"):
+        gr.Markdown("### Interact with the ExperimentAgent")
+
+        with gr.Row():
+            with gr.Column():
+                with gr.Group():
+                    gr.Markdown("### 🌟 Generated Research Idea")
+                    with gr.Row():
+                        idea_input = gr.Textbox(label="Generated Research Hypothesis", lines=30, interactive=False)
+                        plan_input = gr.Textbox(label="Generated Experiment Plan", lines=30, interactive=False)
+        
+            with gr.Column():
+                start_exp_agnet = gr.Button("⚙️ Start / Reset ExperimentAgent", elem_classes=["agent-btn"])
+                with gr.Group():
+                    gr.Markdown("### Implementation + Execution Log")    
+                    log = gr.Textbox(label="📖 Execution Log", lines=20, interactive=False)
+                    code_display = gr.Code(label="🧑‍💻 Implementation", language="python", interactive=False)
+
+            with gr.Column():
+                response = gr.Textbox(label="🤖 ExperimentAgent Response", lines=30, interactive=False)
+                feedback = gr.Textbox(placeholder="N/A", label="🧑‍🔬 User Feedback", lines=3, interactive=True)
+                submit_button = gr.Button("Submit", elem_classes=["Submit-btn"])
+        
+        hypothesis_state.change(
+            fn=load_phase_2_inputs, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[idea_input, plan_input, code_display]
+        )
+        
+        # Start research agent
+        start_exp_agnet.click(
+            fn=start_experiment_agent, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[code_display, log, response, feedback]
+        )
+        
+        submit_button.click(
+            fn=submit_feedback,
+            inputs=[feedback, log, response],
+            outputs=[log, response, code_display, feedback]
+        )
+
+# Test
+if __name__ == "__main__":
+    step_index = 0
+    app.launch(share=True)

.history/app_20250404172217.py

@@ -0,0 +1,333 @@
+import gradio as gr
+from pathlib import Path
+from reactagent.environment import Environment
+from reactagent.agents.agent_research import ResearchAgent
+from reactagent.runner import create_parser
+from reactagent import llm
+from reactagent.users.user import User
+import os
+import json
+
+
+# Global variables to store session state
+env = None
+agent = None
+state_example = False
+state_extract = False
+state_generate = False
+state_agent = False
+state_complete = False
+index_ex = "1"
+
+example_text = [
+    "Research Paper 1: Dataset and Baseline for Automatic Student Feedback Analysis",
+    "Research Paper 2: An Empirical Study on the Impact of Code Review on Software Quality"
+]
+
+# Load example JSON file
+def load_example_data():
+    with open("example/example_data.json", "r") as json_file:
+        example_data = json.load(json_file)
+    
+    for idx in example_data.keys():
+        try:
+            file = example_data[idx]["code_init"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_init"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")
+        try:
+            file = example_data[idx]["code_final"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_final"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")     
+    return example_data
+
+example_data = load_example_data()
+
+# Function to handle the selection of an example and populate the respective fields
+def load_example(example_id):
+    global index_ex 
+    index_ex = str(example_id)
+    example = example_data[index_ex]
+    paper_text = 'Title:\t' +  example['title'] + '\n\nAbstract:\t' + example['abstract']
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return [paper_text, tasks, gaps, keywords, recent_works]
+
+example_text = [load_example(1), load_example(2)]
+
+# Function to handle example clicks
+def load_example_and_set_index(paper_text_input):
+    global index_ex, state_example
+    state_example = True
+    index_ex = str(example_text.index(paper_text_input) + 1)
+    example =  load_example(index_ex)
+
+    return example
+
+
+
+########## Phase 1 ##############
+
+def extract_research_elements(paper_text):
+    global state_extract, index_ex, state_example
+    if not state_example or paper_text == "":
+        return "", "", "", ""
+    state_extract = True
+    if not paper_text.strip().startswith("Title:\t" + example_data[index_ex]["title"]):
+        print("Mismatch detected.")
+        print(paper_text)
+        return "", "", "", ""
+    example = example_data[index_ex]
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return tasks, gaps, keywords, recent_works
+
+
+# Step 2: Generate Research Hypothesis and Experiment Plan
+def generate_and_store(paper_text, tasks, gaps, keywords, recent_works):
+    if (not state_extract or not state_example or paper_text == ""):
+        return "", "", "", ""
+    global state_generate, index_ex
+    state_generate = True
+    hypothesis = example_data[index_ex]['hypothesis']
+    experiment_plan = example_data[index_ex]['experiment_plan']
+    return hypothesis, experiment_plan, hypothesis, experiment_plan
+
+##########  Phase 2 & 3  ##############
+def start_experiment_agent(hypothesis, plan):
+    if (not state_extract or not state_generate or not state_example):
+        return "", "", ""
+    global state_agent, step_index, state_complete
+    state_agent = True
+    step_index = 0
+    state_complete = False
+    # predefined_message = f"Implement the following hypothesis and experiment plan:\n\nHypothesis:\n{hypothesis}\n\nExperiment Plan:\n{plan}"
+    return example_data[index_ex]['code_init'], predefined_action_log, "", ""
+
+def submit_feedback(user_feedback, history, previous_response):
+    if (not state_extract or not  state_generate or not state_agent or not state_example):
+        return "", "", ""
+    global step_index, state_complete
+    step_index += 1
+    msg = history
+    if step_index < len(process_steps):
+        msg += previous_response + "\nUser feedback:" + user_feedback + "\n\n"
+        response_info = process_steps[step_index]
+        response = info_to_message(response_info)  # Convert dictionary to formatted string
+        response += "Please provide feedback based on the history, response entries, and observation, and questions: "
+        step_index += 1
+        msg += response
+    else:
+        state_complete = True
+        response = "Agent Finished."
+        
+    return msg, response, example_data[index_ex]['code_init'] if state_complete else example_data[index_ex]['code_final'], ""
+
+def load_phase_2_inputs(hypothesis, plan):
+    return hypothesis, plan, "# Code implementation will be displayed here after Start ExperimentAgent."
+
+
+
+predefined_action_log = """
+[Reasoning]: To understand the initial structure and functionality of train.py for effective improvements.
+[Action]: Inspect Script (train.py)
+Input: {"script_name": "train.py", "start_line_number": "1", "end_line_number": "74"}
+Objective: Understand the training script, including data processing, [...]
+[Observation]: The train.py script imports [...]. Sets random seeds [...]. Defines [...] Placeholder functions [...] exist without implementation. [...]
+[Feedback]: The script structure is clear, but key functions (train_model, predict) need proper implementation for proposed model training and prediction.\n
+"""
+
+
+predefined_observation = """
+Epoch [1/10],
+Train MSE: 0.543,
+Test MSE: 0.688
+Epoch [2/10],
+Train MSE: 0.242,
+Test MSE: 0.493\n
+"""
+
+# Initialize the global step_index and history
+process_steps = [
+    {
+        "Action": "Inspect Script Lines (train.py)",
+        "Observation": (
+            "The train.py script imports necessary libraries (e.g., pandas, sklearn, torch). "
+            "Sets random seeds for reproducibility. Defines compute_metrics_for_regression function "
+            "to calculate RMSE for different dimensions. Placeholder functions train_model and "
+            "predict exist without implementations."
+        ),
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The script executed successfully. Generated embeddings using the BERT model. Completed "
+            "the training process without errors. Metrics calculation placeholders indicated areas needing implementation."
+        ),
+    },
+    {
+        "Action": "Edit Script (train.py)",
+        "Observation": (
+            "Edited train.py to separate data loading, model definition, training loop, and evaluation into distinct functions. "
+            "The edited train.py now has clearly defined functions"
+            "for data loading (load_data), model definition (build_model), "
+            "training (train_model), and evaluation (evaluate_model). Similarly, eval.py is reorganized to load the model and perform predictions efficiently."
+        ),
+    },
+    {
+        "Action": "Retrieve Model",
+        "Observation": "CNN and BiLSTM retrieved.",
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The model trained over the specified number of epochs. Training and validation loss values are recorded for each epoch, "
+            "the decrease in loss indicates improved model performance."
+        )
+    },
+    {
+        "Action": "Evaluation",
+        "Observation": predefined_observation,
+    }
+]
+def info_to_message(info):
+    msg = ""
+    for k, v in info.items():
+        if isinstance(v, dict):
+            tempv = v
+            v = ""
+            for k2, v2 in tempv.items():
+                v += f"{k2}:\n  {v2}\n"
+        v = User.indent_text(v, 2)
+        msg += '-' * 64
+        msg += '\n'
+        msg += f"{k}:\n{v}\n"
+    return msg
+
+
+def handle_example_click(example_index):
+    global index_ex
+    index_ex = example_index
+    return load_example(index_ex)  # Simply return the text to display it in the textbox
+
+# Gradio Interface
+with gr.Blocks(css=".gr-examples-label {display: none;}", theme=gr.themes.Default()) as app:
+    gr.Markdown("# MLR- Copilot: Machine Learning Research based on LLM Agents")
+    gr.Markdown("###   ")
+    gr.Markdown("## <span style='color:Orange;'> This UI is for predefined example demo only.</span>")
+    gr.Markdown("## <span style='color:Orange;'> To reproduce the results please use [Github](https://github.com/du-nlp-lab/MLR-Copilot/)</span>")
+
+
+
+    gr.Markdown("MLR-Copilot is a framework where LLMs mimic researchers’ thought processes, designed to enhance the productivity of machine learning research by automating the generation and implementation of research ideas. It begins with a research paper, autonomously generating and validating these ideas, while incorporating human feedback to help reach executable research outcomes.")
+
+
+    gr.Markdown("##    ⬇️  Click an example at bottom to start！")
+
+
+    # Use state variables to store generated hypothesis and experiment plan
+    hypothesis_state = gr.State("")
+    experiment_plan_state = gr.State("")
+
+    ########## Phase 1:  Research Idea Generation Tab ##############
+    with gr.Tab("💡Stage 1: Research Idea Generation"):
+        gr.Markdown("### Extract Research Elements and Generate Research Ideas")
+        with gr.Row():
+            with gr.Column():
+                paper_text_input = gr.Textbox(value="", lines=10, label="📑 Research Paper Text", interactive=False)
+
+                # extract_button = gr.Button("🔍 Extract Research Elements")
+                with gr.Row():
+                    tasks_output = gr.Textbox(placeholder="Research task definition", label="Research Tasks", lines=2, interactive=False)
+                    gaps_output = gr.Textbox(placeholder="Research gaps of current works", label="Research Gaps", lines=2, interactive=False)
+                    keywords_output = gr.Textbox(placeholder="Paper keywords", label="Keywords", lines=2, interactive=False)
+                    recent_works_output = gr.Textbox(placeholder="Recent works extracted from Semantic Scholar", label="Recent Works", lines=2, interactive=False)
+            with gr.Column():
+                with gr.Row():  # Move the button to the top 
+                    generate_button = gr.Button("✍️ Generate Research Hypothesis & Experiment Plan")
+                with gr.Group():
+                    gr.Markdown("### 🌟 Research Idea")
+                    with gr.Row():
+                        hypothesis_output = gr.Textbox(label="Generated Hypothesis", lines=20, interactive=False)
+                        experiment_plan_output = gr.Textbox(label="Generated Experiment Plan", lines=20, interactive=False)
+        
+    
+        # Step 1: Extract Research Elements
+        # extract_button.click(
+        #     fn=extract_research_elements, 
+        #     inputs=paper_text_input, 
+        #     outputs=[tasks_output, gaps_output, keywords_output, recent_works_output]
+        # )
+
+        generate_button.click(
+            fn=generate_and_store, 
+            inputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output], 
+            outputs=[hypothesis_output, experiment_plan_output, hypothesis_state, experiment_plan_state]
+        )
+
+        gr.Examples(
+            examples=example_text,
+            inputs=[paper_text_input],
+            outputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output, hypothesis_output, experiment_plan_output],
+            fn=load_example_and_set_index,
+            run_on_click = True,
+            # label="⬇️ Click an example to load"
+        )
+        
+
+
+
+    ########## Phase 2 & 3:  Experiment implementation and execution ##############
+    with gr.Tab("🧪 Stage 2 & Stage 3: Experiment implementation and execution"):
+        gr.Markdown("### Interact with the ExperimentAgent")
+
+        with gr.Row():
+            with gr.Column():
+                with gr.Group():
+                    gr.Markdown("### 🌟 Generated Research Idea")
+                    with gr.Row():
+                        idea_input = gr.Textbox(label="Generated Research Hypothesis", lines=30, interactive=False)
+                        plan_input = gr.Textbox(label="Generated Experiment Plan", lines=30, interactive=False)
+        
+            with gr.Column():
+                start_exp_agnet = gr.Button("⚙️ Start / Reset ExperimentAgent", elem_classes=["agent-btn"])
+                with gr.Group():
+                    gr.Markdown("### Implementation + Execution Log")    
+                    log = gr.Textbox(label="📖 Execution Log", lines=20, interactive=False)
+                    code_display = gr.Code(label="🧑‍💻 Implementation", language="python", interactive=False)
+
+            with gr.Column():
+                response = gr.Textbox(label="🤖 ExperimentAgent Response", lines=30, interactive=False)
+                feedback = gr.Textbox(placeholder="N/A", label="🧑‍🔬 User Feedback", lines=3, interactive=True)
+                submit_button = gr.Button("Submit", elem_classes=["Submit-btn"])
+        
+        hypothesis_state.change(
+            fn=load_phase_2_inputs, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[idea_input, plan_input, code_display]
+        )
+        
+        # Start research agent
+        start_exp_agnet.click(
+            fn=start_experiment_agent, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[code_display, log, response, feedback]
+        )
+        
+        submit_button.click(
+            fn=submit_feedback,
+            inputs=[feedback, log, response],
+            outputs=[log, response, code_display, feedback]
+        )
+
+# Test
+if __name__ == "__main__":
+    step_index = 0
+    app.launch(share=True)

.history/app_20250404172253.py

@@ -0,0 +1,333 @@
+import gradio as gr
+from pathlib import Path
+from reactagent.environment import Environment
+from reactagent.agents.agent_research import ResearchAgent
+from reactagent.runner import create_parser
+from reactagent import llm
+from reactagent.users.user import User
+import os
+import json
+
+
+# Global variables to store session state
+env = None
+agent = None
+state_example = False
+state_extract = False
+state_generate = False
+state_agent = False
+state_complete = False
+index_ex = "1"
+
+example_text = [
+    "Research Paper 1: Dataset and Baseline for Automatic Student Feedback Analysis",
+    "Research Paper 2: An Empirical Study on the Impact of Code Review on Software Quality"
+]
+
+# Load example JSON file
+def load_example_data():
+    with open("example/example_data.json", "r") as json_file:
+        example_data = json.load(json_file)
+    
+    for idx in example_data.keys():
+        try:
+            file = example_data[idx]["code_init"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_init"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")
+        try:
+            file = example_data[idx]["code_final"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_final"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")     
+    return example_data
+
+example_data = load_example_data()
+
+# Function to handle the selection of an example and populate the respective fields
+def load_example(example_id):
+    global index_ex 
+    index_ex = str(example_id)
+    example = example_data[index_ex]
+    paper_text = 'Title:\t' +  example['title'] + '\n\nAbstract:\t' + example['abstract']
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return [paper_text, tasks, gaps, keywords, recent_works]
+
+example_text = [load_example(1), load_example(2)]
+
+# Function to handle example clicks
+def load_example_and_set_index(paper_text_input):
+    global index_ex, state_example
+    state_example = True
+    index_ex = str(example_text.index(paper_text_input) + 1)
+    example =  load_example(index_ex)
+
+    return example
+
+
+
+########## Phase 1 ##############
+
+def extract_research_elements(paper_text):
+    global state_extract, index_ex, state_example
+    if not state_example or paper_text == "":
+        return "", "", "", ""
+    state_extract = True
+    if not paper_text.strip().startswith("Title:\t" + example_data[index_ex]["title"]):
+        print("Mismatch detected.")
+        print(paper_text)
+        return "", "", "", ""
+    example = example_data[index_ex]
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return tasks, gaps, keywords, recent_works
+
+
+# Step 2: Generate Research Hypothesis and Experiment Plan
+def generate_and_store(paper_text, tasks, gaps, keywords, recent_works):
+    if (not state_extract or not state_example or paper_text == ""):
+        return "", "", "", ""
+    global state_generate, index_ex
+    state_generate = True
+    hypothesis = example_data[index_ex]['hypothesis']
+    experiment_plan = example_data[index_ex]['experiment_plan']
+    return hypothesis, experiment_plan, hypothesis, experiment_plan
+
+##########  Phase 2 & 3  ##############
+def start_experiment_agent(hypothesis, plan):
+    if (not state_extract or not state_generate or not state_example):
+        return "", "", ""
+    global state_agent, step_index, state_complete
+    state_agent = True
+    step_index = 0
+    state_complete = False
+    # predefined_message = f"Implement the following hypothesis and experiment plan:\n\nHypothesis:\n{hypothesis}\n\nExperiment Plan:\n{plan}"
+    return example_data[index_ex]['code_init'], predefined_action_log, "", ""
+
+def submit_feedback(user_feedback, history, previous_response):
+    if (not state_extract or not  state_generate or not state_agent or not state_example):
+        return "", "", ""
+    global step_index, state_complete
+    step_index += 1
+    msg = history
+    if step_index < len(process_steps):
+        msg += previous_response + "\nUser feedback:" + user_feedback + "\n\n"
+        response_info = process_steps[step_index]
+        response = info_to_message(response_info)  # Convert dictionary to formatted string
+        response += "Please provide feedback based on the history, response entries, and observation, and questions: "
+        step_index += 1
+        msg += response
+    else:
+        state_complete = True
+        response = "Agent Finished."
+        
+    return msg, response, example_data[index_ex]['code_init'] if state_complete else example_data[index_ex]['code_final'], ""
+
+def load_phase_2_inputs(hypothesis, plan):
+    return hypothesis, plan, "# Code implementation will be displayed here after Start ExperimentAgent."
+
+
+
+predefined_action_log = """
+[Reasoning]: To understand the initial structure and functionality of train.py for effective improvements.
+[Action]: Inspect Script (train.py)
+Input: {"script_name": "train.py", "start_line_number": "1", "end_line_number": "74"}
+Objective: Understand the training script, including data processing, [...]
+[Observation]: The train.py script imports [...]. Sets random seeds [...]. Defines [...] Placeholder functions [...] exist without implementation. [...]
+[Feedback]: The script structure is clear, but key functions (train_model, predict) need proper implementation for proposed model training and prediction.\n
+"""
+
+
+predefined_observation = """
+Epoch [1/10],
+Train MSE: 0.543,
+Test MSE: 0.688
+Epoch [2/10],
+Train MSE: 0.242,
+Test MSE: 0.493\n
+"""
+
+# Initialize the global step_index and history
+process_steps = [
+    {
+        "Action": "Inspect Script Lines (train.py)",
+        "Observation": (
+            "The train.py script imports necessary libraries (e.g., pandas, sklearn, torch). "
+            "Sets random seeds for reproducibility. Defines compute_metrics_for_regression function "
+            "to calculate RMSE for different dimensions. Placeholder functions train_model and "
+            "predict exist without implementations."
+        ),
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The script executed successfully. Generated embeddings using the BERT model. Completed "
+            "the training process without errors. Metrics calculation placeholders indicated areas needing implementation."
+        ),
+    },
+    {
+        "Action": "Edit Script (train.py)",
+        "Observation": (
+            "Edited train.py to separate data loading, model definition, training loop, and evaluation into distinct functions. "
+            "The edited train.py now has clearly defined functions"
+            "for data loading (load_data), model definition (build_model), "
+            "training (train_model), and evaluation (evaluate_model). Similarly, eval.py is reorganized to load the model and perform predictions efficiently."
+        ),
+    },
+    {
+        "Action": "Retrieve Model",
+        "Observation": "CNN and BiLSTM retrieved.",
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The model trained over the specified number of epochs. Training and validation loss values are recorded for each epoch, "
+            "the decrease in loss indicates improved model performance."
+        )
+    },
+    {
+        "Action": "Evaluation",
+        "Observation": predefined_observation,
+    }
+]
+def info_to_message(info):
+    msg = ""
+    for k, v in info.items():
+        if isinstance(v, dict):
+            tempv = v
+            v = ""
+            for k2, v2 in tempv.items():
+                v += f"{k2}:\n  {v2}\n"
+        v = User.indent_text(v, 2)
+        msg += '-' * 64
+        msg += '\n'
+        msg += f"{k}:\n{v}\n"
+    return msg
+
+
+def handle_example_click(example_index):
+    global index_ex
+    index_ex = example_index
+    return load_example(index_ex)  # Simply return the text to display it in the textbox
+
+# Gradio Interface
+with gr.Blocks(css=".gr-examples-label {display: none;}", theme=gr.themes.Default()) as app:
+    gr.Markdown("# MLR- Copilot: Machine Learning Research based on LLM Agents")
+    gr.Markdown("###   ")
+    gr.Markdown("## <span style='color:Orange;'> This UI is for predefined example demo only.</span>")
+    gr.Markdown("## <span style='color:Orange;'> To reproduce the results please use [Github](https://github.com/du-nlp-lab/MLR-Copilot/)</span>")
+
+
+
+    gr.Markdown("MLR-Copilot is a framework where LLMs mimic researchers’ thought processes, designed to enhance the productivity of machine learning research by automating the generation and implementation of research ideas. It begins with a research paper, autonomously generating and validating these ideas, while incorporating human feedback to help reach executable research outcomes.")
+
+
+    gr.Markdown("##    ⬇️  Click an example at bottom to start！")
+
+
+    # Use state variables to store generated hypothesis and experiment plan
+    hypothesis_state = gr.State("")
+    experiment_plan_state = gr.State("")
+
+    ########## Phase 1:  Research Idea Generation Tab ##############
+    with gr.Tab("💡Stage 1: Research Idea Generation"):
+        gr.Markdown("### Extract Research Elements and Generate Research Ideas")
+        with gr.Row():
+            with gr.Column():
+                paper_text_input = gr.Textbox(value="", lines=10, label="📑 Research Paper Text", interactive=False)
+
+                # extract_button = gr.Button("🔍 Extract Research Elements")
+                with gr.Row():
+                    tasks_output = gr.Textbox(placeholder="Research task definition", label="Research Tasks", lines=2, interactive=False)
+                    gaps_output = gr.Textbox(placeholder="Research gaps of current works", label="Research Gaps", lines=2, interactive=False)
+                    keywords_output = gr.Textbox(placeholder="Paper keywords", label="Keywords", lines=2, interactive=False)
+                    recent_works_output = gr.Textbox(placeholder="Recent works extracted from Semantic Scholar", label="Recent Works", lines=2, interactive=False)
+            with gr.Column():
+                with gr.Row():  # Move the button to the top 
+                    generate_button = gr.Button("✍️ Generate Research Hypothesis & Experiment Plan")
+                with gr.Group():
+                    gr.Markdown("### 🌟 Research Idea")
+                    with gr.Row():
+                        hypothesis_output = gr.Textbox(label="Generated Hypothesis", lines=20, interactive=False)
+                        experiment_plan_output = gr.Textbox(label="Generated Experiment Plan", lines=20, interactive=False)
+        
+    
+        # Step 1: Extract Research Elements
+        # extract_button.click(
+        #     fn=extract_research_elements, 
+        #     inputs=paper_text_input, 
+        #     outputs=[tasks_output, gaps_output, keywords_output, recent_works_output]
+        # )
+
+        generate_button.click(
+            fn=generate_and_store, 
+            inputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output], 
+            outputs=[hypothesis_output, experiment_plan_output, hypothesis_state, experiment_plan_state]
+        )
+
+        gr.Examples(
+            examples=example_text,
+            inputs=[paper_text_input],
+            outputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output, hypothesis_output, experiment_plan_output],
+            fn=load_example_and_set_index,
+            run_on_click = True,
+            # label="⬇️ Click an example to load"
+        )
+        
+
+
+
+    ########## Phase 2 & 3:  Experiment implementation and execution ##############
+    with gr.Tab("🧪 Stage 2 & Stage 3: Experiment implementation and execution"):
+        gr.Markdown("### Interact with the ExperimentAgent")
+
+        with gr.Row():
+            with gr.Column():
+                with gr.Group():
+                    gr.Markdown("### 🌟 Generated Research Idea")
+                    with gr.Row():
+                        idea_input = gr.Textbox(label="Generated Research Hypothesis", lines=30, interactive=False)
+                        plan_input = gr.Textbox(label="Generated Experiment Plan", lines=30, interactive=False)
+        
+            with gr.Column():
+                start_exp_agnet = gr.Button("⚙️ Start / Reset ExperimentAgent", elem_classes=["agent-btn"])
+                with gr.Group():
+                    gr.Markdown("### Implementation + Execution Log")    
+                    log = gr.Textbox(label="📖 Execution Log", lines=20, interactive=False)
+                    code_display = gr.Code(label="🧑‍💻 Implementation", language="python", interactive=False)
+
+            with gr.Column():
+                response = gr.Textbox(label="🤖 ExperimentAgent Response", lines=30, interactive=False)
+                feedback = gr.Textbox(placeholder="N/A", label="🧑‍🔬 User Feedback", lines=3, interactive=True)
+                submit_button = gr.Button("Submit", elem_classes=["Submit-btn"])
+        
+        hypothesis_state.change(
+            fn=load_phase_2_inputs, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[idea_input, plan_input, code_display]
+        )
+        
+        # Start research agent
+        start_exp_agnet.click(
+            fn=start_experiment_agent, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[code_display, log, response, feedback]
+        )
+        
+        submit_button.click(
+            fn=submit_feedback,
+            inputs=[feedback, log, response],
+            outputs=[log, response, code_display, feedback]
+        )
+
+# Test
+if __name__ == "__main__":
+    step_index = 0
+    app.launch(share=True)

.history/app_20250404173244.py

@@ -0,0 +1,333 @@
+import gradio as gr
+from pathlib import Path
+from reactagent.environment import Environment
+from reactagent.agents.agent_research import ResearchAgent
+from reactagent.runner import create_parser
+from reactagent import llm
+from reactagent.users.user import User
+import os
+import json
+
+
+# Global variables to store session state
+env = None
+agent = None
+state_example = False
+state_extract = False
+state_generate = False
+state_agent = False
+state_complete = False
+index_ex = "1"
+
+example_text = [
+    "Research Paper 1: Dataset and Baseline for Automatic Student Feedback Analysis",
+    "Research Paper 2: An Empirical Study on the Impact of Code Review on Software Quality"
+]
+
+# Load example JSON file
+def load_example_data():
+    with open("example/example_data.json", "r") as json_file:
+        example_data = json.load(json_file)
+    
+    for idx in example_data.keys():
+        try:
+            file = example_data[idx]["code_init"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_init"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")
+        try:
+            file = example_data[idx]["code_final"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_final"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")     
+    return example_data
+
+example_data = load_example_data()
+
+# Function to handle the selection of an example and populate the respective fields
+def load_example(example_id):
+    global index_ex 
+    index_ex = str(example_id)
+    example = example_data[index_ex]
+    paper_text = 'Title:\t' +  example['title'] + '\n\nAbstract:\t' + example['abstract']
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return paper_text, tasks, gaps, keywords, recent_works
+
+example_text = [load_example(1), load_example(2)]
+
+# Function to handle example clicks
+def load_example_and_set_index(paper_text_input):
+    global index_ex, state_example
+    state_example = True
+    index_ex = str(example_text.index(paper_text_input) + 1)
+    example =  load_example(index_ex)
+
+    return example
+
+
+
+########## Phase 1 ##############
+
+def extract_research_elements(paper_text):
+    global state_extract, index_ex, state_example
+    if not state_example or paper_text == "":
+        return "", "", "", ""
+    state_extract = True
+    if not paper_text.strip().startswith("Title:\t" + example_data[index_ex]["title"]):
+        print("Mismatch detected.")
+        print(paper_text)
+        return "", "", "", ""
+    example = example_data[index_ex]
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return tasks, gaps, keywords, recent_works
+
+
+# Step 2: Generate Research Hypothesis and Experiment Plan
+def generate_and_store(paper_text, tasks, gaps, keywords, recent_works):
+    if (not state_extract or not state_example or paper_text == ""):
+        return "", "", "", ""
+    global state_generate, index_ex
+    state_generate = True
+    hypothesis = example_data[index_ex]['hypothesis']
+    experiment_plan = example_data[index_ex]['experiment_plan']
+    return hypothesis, experiment_plan, hypothesis, experiment_plan
+
+##########  Phase 2 & 3  ##############
+def start_experiment_agent(hypothesis, plan):
+    if (not state_extract or not state_generate or not state_example):
+        return "", "", ""
+    global state_agent, step_index, state_complete
+    state_agent = True
+    step_index = 0
+    state_complete = False
+    # predefined_message = f"Implement the following hypothesis and experiment plan:\n\nHypothesis:\n{hypothesis}\n\nExperiment Plan:\n{plan}"
+    return example_data[index_ex]['code_init'], predefined_action_log, "", ""
+
+def submit_feedback(user_feedback, history, previous_response):
+    if (not state_extract or not  state_generate or not state_agent or not state_example):
+        return "", "", ""
+    global step_index, state_complete
+    step_index += 1
+    msg = history
+    if step_index < len(process_steps):
+        msg += previous_response + "\nUser feedback:" + user_feedback + "\n\n"
+        response_info = process_steps[step_index]
+        response = info_to_message(response_info)  # Convert dictionary to formatted string
+        response += "Please provide feedback based on the history, response entries, and observation, and questions: "
+        step_index += 1
+        msg += response
+    else:
+        state_complete = True
+        response = "Agent Finished."
+        
+    return msg, response, example_data[index_ex]['code_init'] if state_complete else example_data[index_ex]['code_final'], ""
+
+def load_phase_2_inputs(hypothesis, plan):
+    return hypothesis, plan, "# Code implementation will be displayed here after Start ExperimentAgent."
+
+
+
+predefined_action_log = """
+[Reasoning]: To understand the initial structure and functionality of train.py for effective improvements.
+[Action]: Inspect Script (train.py)
+Input: {"script_name": "train.py", "start_line_number": "1", "end_line_number": "74"}
+Objective: Understand the training script, including data processing, [...]
+[Observation]: The train.py script imports [...]. Sets random seeds [...]. Defines [...] Placeholder functions [...] exist without implementation. [...]
+[Feedback]: The script structure is clear, but key functions (train_model, predict) need proper implementation for proposed model training and prediction.\n
+"""
+
+
+predefined_observation = """
+Epoch [1/10],
+Train MSE: 0.543,
+Test MSE: 0.688
+Epoch [2/10],
+Train MSE: 0.242,
+Test MSE: 0.493\n
+"""
+
+# Initialize the global step_index and history
+process_steps = [
+    {
+        "Action": "Inspect Script Lines (train.py)",
+        "Observation": (
+            "The train.py script imports necessary libraries (e.g., pandas, sklearn, torch). "
+            "Sets random seeds for reproducibility. Defines compute_metrics_for_regression function "
+            "to calculate RMSE for different dimensions. Placeholder functions train_model and "
+            "predict exist without implementations."
+        ),
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The script executed successfully. Generated embeddings using the BERT model. Completed "
+            "the training process without errors. Metrics calculation placeholders indicated areas needing implementation."
+        ),
+    },
+    {
+        "Action": "Edit Script (train.py)",
+        "Observation": (
+            "Edited train.py to separate data loading, model definition, training loop, and evaluation into distinct functions. "
+            "The edited train.py now has clearly defined functions"
+            "for data loading (load_data), model definition (build_model), "
+            "training (train_model), and evaluation (evaluate_model). Similarly, eval.py is reorganized to load the model and perform predictions efficiently."
+        ),
+    },
+    {
+        "Action": "Retrieve Model",
+        "Observation": "CNN and BiLSTM retrieved.",
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The model trained over the specified number of epochs. Training and validation loss values are recorded for each epoch, "
+            "the decrease in loss indicates improved model performance."
+        )
+    },
+    {
+        "Action": "Evaluation",
+        "Observation": predefined_observation,
+    }
+]
+def info_to_message(info):
+    msg = ""
+    for k, v in info.items():
+        if isinstance(v, dict):
+            tempv = v
+            v = ""
+            for k2, v2 in tempv.items():
+                v += f"{k2}:\n  {v2}\n"
+        v = User.indent_text(v, 2)
+        msg += '-' * 64
+        msg += '\n'
+        msg += f"{k}:\n{v}\n"
+    return msg
+
+
+def handle_example_click(example_index):
+    global index_ex
+    index_ex = example_index
+    return load_example(index_ex)  # Simply return the text to display it in the textbox
+
+# Gradio Interface
+with gr.Blocks(css=".gr-examples-label {display: none;}", theme=gr.themes.Default()) as app:
+    gr.Markdown("# MLR- Copilot: Machine Learning Research based on LLM Agents")
+    gr.Markdown("###   ")
+    gr.Markdown("## <span style='color:Orange;'> This UI is for predefined example demo only.</span>")
+    gr.Markdown("## <span style='color:Orange;'> To reproduce the results please use [Github](https://github.com/du-nlp-lab/MLR-Copilot/)</span>")
+
+
+
+    gr.Markdown("MLR-Copilot is a framework where LLMs mimic researchers’ thought processes, designed to enhance the productivity of machine learning research by automating the generation and implementation of research ideas. It begins with a research paper, autonomously generating and validating these ideas, while incorporating human feedback to help reach executable research outcomes.")
+
+
+    gr.Markdown("##    ⬇️  Click an example at bottom to start！")
+
+
+    # Use state variables to store generated hypothesis and experiment plan
+    hypothesis_state = gr.State("")
+    experiment_plan_state = gr.State("")
+
+    ########## Phase 1:  Research Idea Generation Tab ##############
+    with gr.Tab("💡Stage 1: Research Idea Generation"):
+        gr.Markdown("### Extract Research Elements and Generate Research Ideas")
+        with gr.Row():
+            with gr.Column():
+                paper_text_input = gr.Textbox(value="", lines=10, label="📑 Research Paper Text", interactive=False)
+
+                # extract_button = gr.Button("🔍 Extract Research Elements")
+                with gr.Row():
+                    tasks_output = gr.Textbox(placeholder="Research task definition", label="Research Tasks", lines=2, interactive=False)
+                    gaps_output = gr.Textbox(placeholder="Research gaps of current works", label="Research Gaps", lines=2, interactive=False)
+                    keywords_output = gr.Textbox(placeholder="Paper keywords", label="Keywords", lines=2, interactive=False)
+                    recent_works_output = gr.Textbox(placeholder="Recent works extracted from Semantic Scholar", label="Recent Works", lines=2, interactive=False)
+            with gr.Column():
+                with gr.Row():  # Move the button to the top 
+                    generate_button = gr.Button("✍️ Generate Research Hypothesis & Experiment Plan")
+                with gr.Group():
+                    gr.Markdown("### 🌟 Research Idea")
+                    with gr.Row():
+                        hypothesis_output = gr.Textbox(label="Generated Hypothesis", lines=20, interactive=False)
+                        experiment_plan_output = gr.Textbox(label="Generated Experiment Plan", lines=20, interactive=False)
+        
+    
+        # Step 1: Extract Research Elements
+        # extract_button.click(
+        #     fn=extract_research_elements, 
+        #     inputs=paper_text_input, 
+        #     outputs=[tasks_output, gaps_output, keywords_output, recent_works_output]
+        # )
+
+        generate_button.click(
+            fn=generate_and_store, 
+            inputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output], 
+            outputs=[hypothesis_output, experiment_plan_output, hypothesis_state, experiment_plan_state]
+        )
+
+        gr.Examples(
+            examples=example_text,
+            inputs=[paper_text_input],
+            outputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output, hypothesis_output, experiment_plan_output],
+            fn=load_example_and_set_index,
+            run_on_click = True,
+            # label="⬇️ Click an example to load"
+        )
+        
+
+
+
+    ########## Phase 2 & 3:  Experiment implementation and execution ##############
+    with gr.Tab("🧪 Stage 2 & Stage 3: Experiment implementation and execution"):
+        gr.Markdown("### Interact with the ExperimentAgent")
+
+        with gr.Row():
+            with gr.Column():
+                with gr.Group():
+                    gr.Markdown("### 🌟 Generated Research Idea")
+                    with gr.Row():
+                        idea_input = gr.Textbox(label="Generated Research Hypothesis", lines=30, interactive=False)
+                        plan_input = gr.Textbox(label="Generated Experiment Plan", lines=30, interactive=False)
+        
+            with gr.Column():
+                start_exp_agnet = gr.Button("⚙️ Start / Reset ExperimentAgent", elem_classes=["agent-btn"])
+                with gr.Group():
+                    gr.Markdown("### Implementation + Execution Log")    
+                    log = gr.Textbox(label="📖 Execution Log", lines=20, interactive=False)
+                    code_display = gr.Code(label="🧑‍💻 Implementation", language="python", interactive=False)
+
+            with gr.Column():
+                response = gr.Textbox(label="🤖 ExperimentAgent Response", lines=30, interactive=False)
+                feedback = gr.Textbox(placeholder="N/A", label="🧑‍🔬 User Feedback", lines=3, interactive=True)
+                submit_button = gr.Button("Submit", elem_classes=["Submit-btn"])
+        
+        hypothesis_state.change(
+            fn=load_phase_2_inputs, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[idea_input, plan_input, code_display]
+        )
+        
+        # Start research agent
+        start_exp_agnet.click(
+            fn=start_experiment_agent, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[code_display, log, response, feedback]
+        )
+        
+        submit_button.click(
+            fn=submit_feedback,
+            inputs=[feedback, log, response],
+            outputs=[log, response, code_display, feedback]
+        )
+
+# Test
+if __name__ == "__main__":
+    step_index = 0
+    app.launch(share=True)

.history/app_20250404173256.py

@@ -0,0 +1,333 @@
+import gradio as gr
+from pathlib import Path
+from reactagent.environment import Environment
+from reactagent.agents.agent_research import ResearchAgent
+from reactagent.runner import create_parser
+from reactagent import llm
+from reactagent.users.user import User
+import os
+import json
+
+
+# Global variables to store session state
+env = None
+agent = None
+state_example = False
+state_extract = False
+state_generate = False
+state_agent = False
+state_complete = False
+index_ex = "1"
+
+example_text = [
+    "Research Paper 1: Dataset and Baseline for Automatic Student Feedback Analysis",
+    "Research Paper 2: An Empirical Study on the Impact of Code Review on Software Quality"
+]
+
+# Load example JSON file
+def load_example_data():
+    with open("example/example_data.json", "r") as json_file:
+        example_data = json.load(json_file)
+    
+    for idx in example_data.keys():
+        try:
+            file = example_data[idx]["code_init"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_init"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")
+        try:
+            file = example_data[idx]["code_final"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_final"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")     
+    return example_data
+
+example_data = load_example_data()
+
+# Function to handle the selection of an example and populate the respective fields
+def load_example(example_id):
+    global index_ex 
+    index_ex = str(example_id)
+    example = example_data[index_ex]
+    paper_text = 'Title:\t' +  example['title'] + '\n\nAbstract:\t' + example['abstract']
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return [paper_text, tasks, gaps, keywords, recent_works]
+
+example_text = [load_example(1), load_example(2)]
+
+# Function to handle example clicks
+def load_example_and_set_index(paper_text_input):
+    global index_ex, state_example
+    state_example = True
+    index_ex = str(example_text.index(paper_text_input) + 1)
+    example =  load_example(index_ex)
+
+    return example
+
+
+
+########## Phase 1 ##############
+
+def extract_research_elements(paper_text):
+    global state_extract, index_ex, state_example
+    if not state_example or paper_text == "":
+        return "", "", "", ""
+    state_extract = True
+    if not paper_text.strip().startswith("Title:\t" + example_data[index_ex]["title"]):
+        print("Mismatch detected.")
+        print(paper_text)
+        return "", "", "", ""
+    example = example_data[index_ex]
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return tasks, gaps, keywords, recent_works
+
+
+# Step 2: Generate Research Hypothesis and Experiment Plan
+def generate_and_store(paper_text, tasks, gaps, keywords, recent_works):
+    if (not state_extract or not state_example or paper_text == ""):
+        return "", "", "", ""
+    global state_generate, index_ex
+    state_generate = True
+    hypothesis = example_data[index_ex]['hypothesis']
+    experiment_plan = example_data[index_ex]['experiment_plan']
+    return hypothesis, experiment_plan, hypothesis, experiment_plan
+
+##########  Phase 2 & 3  ##############
+def start_experiment_agent(hypothesis, plan):
+    if (not state_extract or not state_generate or not state_example):
+        return "", "", ""
+    global state_agent, step_index, state_complete
+    state_agent = True
+    step_index = 0
+    state_complete = False
+    # predefined_message = f"Implement the following hypothesis and experiment plan:\n\nHypothesis:\n{hypothesis}\n\nExperiment Plan:\n{plan}"
+    return example_data[index_ex]['code_init'], predefined_action_log, "", ""
+
+def submit_feedback(user_feedback, history, previous_response):
+    if (not state_extract or not  state_generate or not state_agent or not state_example):
+        return "", "", ""
+    global step_index, state_complete
+    step_index += 1
+    msg = history
+    if step_index < len(process_steps):
+        msg += previous_response + "\nUser feedback:" + user_feedback + "\n\n"
+        response_info = process_steps[step_index]
+        response = info_to_message(response_info)  # Convert dictionary to formatted string
+        response += "Please provide feedback based on the history, response entries, and observation, and questions: "
+        step_index += 1
+        msg += response
+    else:
+        state_complete = True
+        response = "Agent Finished."
+        
+    return msg, response, example_data[index_ex]['code_init'] if state_complete else example_data[index_ex]['code_final'], ""
+
+def load_phase_2_inputs(hypothesis, plan):
+    return hypothesis, plan, "# Code implementation will be displayed here after Start ExperimentAgent."
+
+
+
+predefined_action_log = """
+[Reasoning]: To understand the initial structure and functionality of train.py for effective improvements.
+[Action]: Inspect Script (train.py)
+Input: {"script_name": "train.py", "start_line_number": "1", "end_line_number": "74"}
+Objective: Understand the training script, including data processing, [...]
+[Observation]: The train.py script imports [...]. Sets random seeds [...]. Defines [...] Placeholder functions [...] exist without implementation. [...]
+[Feedback]: The script structure is clear, but key functions (train_model, predict) need proper implementation for proposed model training and prediction.\n
+"""
+
+
+predefined_observation = """
+Epoch [1/10],
+Train MSE: 0.543,
+Test MSE: 0.688
+Epoch [2/10],
+Train MSE: 0.242,
+Test MSE: 0.493\n
+"""
+
+# Initialize the global step_index and history
+process_steps = [
+    {
+        "Action": "Inspect Script Lines (train.py)",
+        "Observation": (
+            "The train.py script imports necessary libraries (e.g., pandas, sklearn, torch). "
+            "Sets random seeds for reproducibility. Defines compute_metrics_for_regression function "
+            "to calculate RMSE for different dimensions. Placeholder functions train_model and "
+            "predict exist without implementations."
+        ),
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The script executed successfully. Generated embeddings using the BERT model. Completed "
+            "the training process without errors. Metrics calculation placeholders indicated areas needing implementation."
+        ),
+    },
+    {
+        "Action": "Edit Script (train.py)",
+        "Observation": (
+            "Edited train.py to separate data loading, model definition, training loop, and evaluation into distinct functions. "
+            "The edited train.py now has clearly defined functions"
+            "for data loading (load_data), model definition (build_model), "
+            "training (train_model), and evaluation (evaluate_model). Similarly, eval.py is reorganized to load the model and perform predictions efficiently."
+        ),
+    },
+    {
+        "Action": "Retrieve Model",
+        "Observation": "CNN and BiLSTM retrieved.",
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The model trained over the specified number of epochs. Training and validation loss values are recorded for each epoch, "
+            "the decrease in loss indicates improved model performance."
+        )
+    },
+    {
+        "Action": "Evaluation",
+        "Observation": predefined_observation,
+    }
+]
+def info_to_message(info):
+    msg = ""
+    for k, v in info.items():
+        if isinstance(v, dict):
+            tempv = v
+            v = ""
+            for k2, v2 in tempv.items():
+                v += f"{k2}:\n  {v2}\n"
+        v = User.indent_text(v, 2)
+        msg += '-' * 64
+        msg += '\n'
+        msg += f"{k}:\n{v}\n"
+    return msg
+
+
+def handle_example_click(example_index):
+    global index_ex
+    index_ex = example_index
+    return load_example(index_ex)  # Simply return the text to display it in the textbox
+
+# Gradio Interface
+with gr.Blocks(css=".gr-examples-label {display: none;}", theme=gr.themes.Default()) as app:
+    gr.Markdown("# MLR- Copilot: Machine Learning Research based on LLM Agents")
+    gr.Markdown("###   ")
+    gr.Markdown("## <span style='color:Orange;'> This UI is for predefined example demo only.</span>")
+    gr.Markdown("## <span style='color:Orange;'> To reproduce the results please use [Github](https://github.com/du-nlp-lab/MLR-Copilot/)</span>")
+
+
+
+    gr.Markdown("MLR-Copilot is a framework where LLMs mimic researchers’ thought processes, designed to enhance the productivity of machine learning research by automating the generation and implementation of research ideas. It begins with a research paper, autonomously generating and validating these ideas, while incorporating human feedback to help reach executable research outcomes.")
+
+
+    gr.Markdown("##    ⬇️  Click an example at bottom to start！")
+
+
+    # Use state variables to store generated hypothesis and experiment plan
+    hypothesis_state = gr.State("")
+    experiment_plan_state = gr.State("")
+
+    ########## Phase 1:  Research Idea Generation Tab ##############
+    with gr.Tab("💡Stage 1: Research Idea Generation"):
+        gr.Markdown("### Extract Research Elements and Generate Research Ideas")
+        with gr.Row():
+            with gr.Column():
+                paper_text_input = gr.Textbox(value="", lines=10, label="📑 Research Paper Text", interactive=False)
+
+                # extract_button = gr.Button("🔍 Extract Research Elements")
+                with gr.Row():
+                    tasks_output = gr.Textbox(placeholder="Research task definition", label="Research Tasks", lines=2, interactive=False)
+                    gaps_output = gr.Textbox(placeholder="Research gaps of current works", label="Research Gaps", lines=2, interactive=False)
+                    keywords_output = gr.Textbox(placeholder="Paper keywords", label="Keywords", lines=2, interactive=False)
+                    recent_works_output = gr.Textbox(placeholder="Recent works extracted from Semantic Scholar", label="Recent Works", lines=2, interactive=False)
+            with gr.Column():
+                with gr.Row():  # Move the button to the top 
+                    generate_button = gr.Button("✍️ Generate Research Hypothesis & Experiment Plan")
+                with gr.Group():
+                    gr.Markdown("### 🌟 Research Idea")
+                    with gr.Row():
+                        hypothesis_output = gr.Textbox(label="Generated Hypothesis", lines=20, interactive=False)
+                        experiment_plan_output = gr.Textbox(label="Generated Experiment Plan", lines=20, interactive=False)
+        
+    
+        # Step 1: Extract Research Elements
+        # extract_button.click(
+        #     fn=extract_research_elements, 
+        #     inputs=paper_text_input, 
+        #     outputs=[tasks_output, gaps_output, keywords_output, recent_works_output]
+        # )
+
+        generate_button.click(
+            fn=generate_and_store, 
+            inputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output], 
+            outputs=[hypothesis_output, experiment_plan_output, hypothesis_state, experiment_plan_state]
+        )
+
+        gr.Examples(
+            examples=example_text,
+            inputs=[paper_text_input],
+            outputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output, hypothesis_output, experiment_plan_output],
+            fn=load_example_and_set_index,
+            run_on_click = True,
+            # label="⬇️ Click an example to load"
+        )
+        
+
+
+
+    ########## Phase 2 & 3:  Experiment implementation and execution ##############
+    with gr.Tab("🧪 Stage 2 & Stage 3: Experiment implementation and execution"):
+        gr.Markdown("### Interact with the ExperimentAgent")
+
+        with gr.Row():
+            with gr.Column():
+                with gr.Group():
+                    gr.Markdown("### 🌟 Generated Research Idea")
+                    with gr.Row():
+                        idea_input = gr.Textbox(label="Generated Research Hypothesis", lines=30, interactive=False)
+                        plan_input = gr.Textbox(label="Generated Experiment Plan", lines=30, interactive=False)
+        
+            with gr.Column():
+                start_exp_agnet = gr.Button("⚙️ Start / Reset ExperimentAgent", elem_classes=["agent-btn"])
+                with gr.Group():
+                    gr.Markdown("### Implementation + Execution Log")    
+                    log = gr.Textbox(label="📖 Execution Log", lines=20, interactive=False)
+                    code_display = gr.Code(label="🧑‍💻 Implementation", language="python", interactive=False)
+
+            with gr.Column():
+                response = gr.Textbox(label="🤖 ExperimentAgent Response", lines=30, interactive=False)
+                feedback = gr.Textbox(placeholder="N/A", label="🧑‍🔬 User Feedback", lines=3, interactive=True)
+                submit_button = gr.Button("Submit", elem_classes=["Submit-btn"])
+        
+        hypothesis_state.change(
+            fn=load_phase_2_inputs, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[idea_input, plan_input, code_display]
+        )
+        
+        # Start research agent
+        start_exp_agnet.click(
+            fn=start_experiment_agent, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[code_display, log, response, feedback]
+        )
+        
+        submit_button.click(
+            fn=submit_feedback,
+            inputs=[feedback, log, response],
+            outputs=[log, response, code_display, feedback]
+        )
+
+# Test
+if __name__ == "__main__":
+    step_index = 0
+    app.launch(share=True)

.history/app_20250404173257.py

@@ -0,0 +1,333 @@
+import gradio as gr
+from pathlib import Path
+from reactagent.environment import Environment
+from reactagent.agents.agent_research import ResearchAgent
+from reactagent.runner import create_parser
+from reactagent import llm
+from reactagent.users.user import User
+import os
+import json
+
+
+# Global variables to store session state
+env = None
+agent = None
+state_example = False
+state_extract = False
+state_generate = False
+state_agent = False
+state_complete = False
+index_ex = "1"
+
+example_text = [
+    "Research Paper 1: Dataset and Baseline for Automatic Student Feedback Analysis",
+    "Research Paper 2: An Empirical Study on the Impact of Code Review on Software Quality"
+]
+
+# Load example JSON file
+def load_example_data():
+    with open("example/example_data.json", "r") as json_file:
+        example_data = json.load(json_file)
+    
+    for idx in example_data.keys():
+        try:
+            file = example_data[idx]["code_init"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_init"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")
+        try:
+            file = example_data[idx]["code_final"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_final"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")     
+    return example_data
+
+example_data = load_example_data()
+
+# Function to handle the selection of an example and populate the respective fields
+def load_example(example_id):
+    global index_ex 
+    index_ex = str(example_id)
+    example = example_data[index_ex]
+    paper_text = 'Title:\t' +  example['title'] + '\n\nAbstract:\t' + example['abstract']
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return [paper_text, tasks, gaps, keywords, recent_works]
+
+example_text = [load_example(1), load_example(2)]
+
+# Function to handle example clicks
+def load_example_and_set_index(paper_text_input):
+    global index_ex, state_example
+    state_example = True
+    index_ex = str(example_text.index(paper_text_input) + 1)
+    example =  load_example(index_ex)
+
+    return example
+
+
+
+########## Phase 1 ##############
+
+def extract_research_elements(paper_text):
+    global state_extract, index_ex, state_example
+    if not state_example or paper_text == "":
+        return "", "", "", ""
+    state_extract = True
+    if not paper_text.strip().startswith("Title:\t" + example_data[index_ex]["title"]):
+        print("Mismatch detected.")
+        print(paper_text)
+        return "", "", "", ""
+    example = example_data[index_ex]
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return tasks, gaps, keywords, recent_works
+
+
+# Step 2: Generate Research Hypothesis and Experiment Plan
+def generate_and_store(paper_text, tasks, gaps, keywords, recent_works):
+    if (not state_extract or not state_example or paper_text == ""):
+        return "", "", "", ""
+    global state_generate, index_ex
+    state_generate = True
+    hypothesis = example_data[index_ex]['hypothesis']
+    experiment_plan = example_data[index_ex]['experiment_plan']
+    return hypothesis, experiment_plan, hypothesis, experiment_plan
+
+##########  Phase 2 & 3  ##############
+def start_experiment_agent(hypothesis, plan):
+    if (not state_extract or not state_generate or not state_example):
+        return "", "", ""
+    global state_agent, step_index, state_complete
+    state_agent = True
+    step_index = 0
+    state_complete = False
+    # predefined_message = f"Implement the following hypothesis and experiment plan:\n\nHypothesis:\n{hypothesis}\n\nExperiment Plan:\n{plan}"
+    return example_data[index_ex]['code_init'], predefined_action_log, "", ""
+
+def submit_feedback(user_feedback, history, previous_response):
+    if (not state_extract or not  state_generate or not state_agent or not state_example):
+        return "", "", ""
+    global step_index, state_complete
+    step_index += 1
+    msg = history
+    if step_index < len(process_steps):
+        msg += previous_response + "\nUser feedback:" + user_feedback + "\n\n"
+        response_info = process_steps[step_index]
+        response = info_to_message(response_info)  # Convert dictionary to formatted string
+        response += "Please provide feedback based on the history, response entries, and observation, and questions: "
+        step_index += 1
+        msg += response
+    else:
+        state_complete = True
+        response = "Agent Finished."
+        
+    return msg, response, example_data[index_ex]['code_init'] if state_complete else example_data[index_ex]['code_final'], ""
+
+def load_phase_2_inputs(hypothesis, plan):
+    return hypothesis, plan, "# Code implementation will be displayed here after Start ExperimentAgent."
+
+
+
+predefined_action_log = """
+[Reasoning]: To understand the initial structure and functionality of train.py for effective improvements.
+[Action]: Inspect Script (train.py)
+Input: {"script_name": "train.py", "start_line_number": "1", "end_line_number": "74"}
+Objective: Understand the training script, including data processing, [...]
+[Observation]: The train.py script imports [...]. Sets random seeds [...]. Defines [...] Placeholder functions [...] exist without implementation. [...]
+[Feedback]: The script structure is clear, but key functions (train_model, predict) need proper implementation for proposed model training and prediction.\n
+"""
+
+
+predefined_observation = """
+Epoch [1/10],
+Train MSE: 0.543,
+Test MSE: 0.688
+Epoch [2/10],
+Train MSE: 0.242,
+Test MSE: 0.493\n
+"""
+
+# Initialize the global step_index and history
+process_steps = [
+    {
+        "Action": "Inspect Script Lines (train.py)",
+        "Observation": (
+            "The train.py script imports necessary libraries (e.g., pandas, sklearn, torch). "
+            "Sets random seeds for reproducibility. Defines compute_metrics_for_regression function "
+            "to calculate RMSE for different dimensions. Placeholder functions train_model and "
+            "predict exist without implementations."
+        ),
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The script executed successfully. Generated embeddings using the BERT model. Completed "
+            "the training process without errors. Metrics calculation placeholders indicated areas needing implementation."
+        ),
+    },
+    {
+        "Action": "Edit Script (train.py)",
+        "Observation": (
+            "Edited train.py to separate data loading, model definition, training loop, and evaluation into distinct functions. "
+            "The edited train.py now has clearly defined functions"
+            "for data loading (load_data), model definition (build_model), "
+            "training (train_model), and evaluation (evaluate_model). Similarly, eval.py is reorganized to load the model and perform predictions efficiently."
+        ),
+    },
+    {
+        "Action": "Retrieve Model",
+        "Observation": "CNN and BiLSTM retrieved.",
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The model trained over the specified number of epochs. Training and validation loss values are recorded for each epoch, "
+            "the decrease in loss indicates improved model performance."
+        )
+    },
+    {
+        "Action": "Evaluation",
+        "Observation": predefined_observation,
+    }
+]
+def info_to_message(info):
+    msg = ""
+    for k, v in info.items():
+        if isinstance(v, dict):
+            tempv = v
+            v = ""
+            for k2, v2 in tempv.items():
+                v += f"{k2}:\n  {v2}\n"
+        v = User.indent_text(v, 2)
+        msg += '-' * 64
+        msg += '\n'
+        msg += f"{k}:\n{v}\n"
+    return msg
+
+
+def handle_example_click(example_index):
+    global index_ex
+    index_ex = example_index
+    return load_example(index_ex)  # Simply return the text to display it in the textbox
+
+# Gradio Interface
+with gr.Blocks(css=".gr-examples-label {display: none;}", theme=gr.themes.Default()) as app:
+    gr.Markdown("# MLR- Copilot: Machine Learning Research based on LLM Agents")
+    gr.Markdown("###   ")
+    gr.Markdown("## <span style='color:Orange;'> This UI is for predefined example demo only.</span>")
+    gr.Markdown("## <span style='color:Orange;'> To reproduce the results please use [Github](https://github.com/du-nlp-lab/MLR-Copilot/)</span>")
+
+
+
+    gr.Markdown("MLR-Copilot is a framework where LLMs mimic researchers’ thought processes, designed to enhance the productivity of machine learning research by automating the generation and implementation of research ideas. It begins with a research paper, autonomously generating and validating these ideas, while incorporating human feedback to help reach executable research outcomes.")
+
+
+    gr.Markdown("##    ⬇️  Click an example at bottom to start！")
+
+
+    # Use state variables to store generated hypothesis and experiment plan
+    hypothesis_state = gr.State("")
+    experiment_plan_state = gr.State("")
+
+    ########## Phase 1:  Research Idea Generation Tab ##############
+    with gr.Tab("💡Stage 1: Research Idea Generation"):
+        gr.Markdown("### Extract Research Elements and Generate Research Ideas")
+        with gr.Row():
+            with gr.Column():
+                paper_text_input = gr.Textbox(value="", lines=10, label="📑 Research Paper Text", interactive=False)
+
+                # extract_button = gr.Button("🔍 Extract Research Elements")
+                with gr.Row():
+                    tasks_output = gr.Textbox(placeholder="Research task definition", label="Research Tasks", lines=2, interactive=False)
+                    gaps_output = gr.Textbox(placeholder="Research gaps of current works", label="Research Gaps", lines=2, interactive=False)
+                    keywords_output = gr.Textbox(placeholder="Paper keywords", label="Keywords", lines=2, interactive=False)
+                    recent_works_output = gr.Textbox(placeholder="Recent works extracted from Semantic Scholar", label="Recent Works", lines=2, interactive=False)
+            with gr.Column():
+                with gr.Row():  # Move the button to the top 
+                    generate_button = gr.Button("✍️ Generate Research Hypothesis & Experiment Plan")
+                with gr.Group():
+                    gr.Markdown("### 🌟 Research Idea")
+                    with gr.Row():
+                        hypothesis_output = gr.Textbox(label="Generated Hypothesis", lines=20, interactive=False)
+                        experiment_plan_output = gr.Textbox(label="Generated Experiment Plan", lines=20, interactive=False)
+        
+    
+        # Step 1: Extract Research Elements
+        # extract_button.click(
+        #     fn=extract_research_elements, 
+        #     inputs=paper_text_input, 
+        #     outputs=[tasks_output, gaps_output, keywords_output, recent_works_output]
+        # )
+
+        generate_button.click(
+            fn=generate_and_store, 
+            inputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output], 
+            outputs=[hypothesis_output, experiment_plan_output, hypothesis_state, experiment_plan_state]
+        )
+
+        gr.Examples(
+            examples=example_text,
+            inputs=[paper_text_input],
+            outputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output, hypothesis_output, experiment_plan_output],
+            fn=load_example_and_set_index,
+            run_on_click = True,
+            # label="⬇️ Click an example to load"
+        )
+        
+
+
+
+    ########## Phase 2 & 3:  Experiment implementation and execution ##############
+    with gr.Tab("🧪 Stage 2 & Stage 3: Experiment implementation and execution"):
+        gr.Markdown("### Interact with the ExperimentAgent")
+
+        with gr.Row():
+            with gr.Column():
+                with gr.Group():
+                    gr.Markdown("### 🌟 Generated Research Idea")
+                    with gr.Row():
+                        idea_input = gr.Textbox(label="Generated Research Hypothesis", lines=30, interactive=False)
+                        plan_input = gr.Textbox(label="Generated Experiment Plan", lines=30, interactive=False)
+        
+            with gr.Column():
+                start_exp_agnet = gr.Button("⚙️ Start / Reset ExperimentAgent", elem_classes=["agent-btn"])
+                with gr.Group():
+                    gr.Markdown("### Implementation + Execution Log")    
+                    log = gr.Textbox(label="📖 Execution Log", lines=20, interactive=False)
+                    code_display = gr.Code(label="🧑‍💻 Implementation", language="python", interactive=False)
+
+            with gr.Column():
+                response = gr.Textbox(label="🤖 ExperimentAgent Response", lines=30, interactive=False)
+                feedback = gr.Textbox(placeholder="N/A", label="🧑‍🔬 User Feedback", lines=3, interactive=True)
+                submit_button = gr.Button("Submit", elem_classes=["Submit-btn"])
+        
+        hypothesis_state.change(
+            fn=load_phase_2_inputs, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[idea_input, plan_input, code_display]
+        )
+        
+        # Start research agent
+        start_exp_agnet.click(
+            fn=start_experiment_agent, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[code_display, log, response, feedback]
+        )
+        
+        submit_button.click(
+            fn=submit_feedback,
+            inputs=[feedback, log, response],
+            outputs=[log, response, code_display, feedback]
+        )
+
+# Test
+if __name__ == "__main__":
+    step_index = 0
+    app.launch(share=True)

.history/app_20250404174324.py

@@ -0,0 +1,341 @@
+import gradio as gr
+from pathlib import Path
+from reactagent.environment import Environment
+from reactagent.agents.agent_research import ResearchAgent
+from reactagent.runner import create_parser
+from reactagent import llm
+from reactagent.users.user import User
+import os
+import json
+
+
+# Global variables to store session state
+env = None
+agent = None
+state_example = False
+state_extract = False
+state_generate = False
+state_agent = False
+state_complete = False
+index_ex = "1"
+
+example_text = [
+    "Research Paper 1: Dataset and Baseline for Automatic Student Feedback Analysis",
+    "Research Paper 2: An Empirical Study on the Impact of Code Review on Software Quality"
+]
+
+
+# Load example JSON file
+def load_example_data():
+    with open("example/example_data.json", "r") as json_file:
+        example_data = json.load(json_file)
+    
+    for idx in example_data.keys():
+        try:
+            file = example_data[idx]["code_init"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_init"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")
+        try:
+            file = example_data[idx]["code_final"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_final"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")     
+    return example_data
+
+example_data = load_example_data()
+
+# Function to handle the selection of an example and populate the respective fields
+def load_example(example_id):
+    global index_ex 
+    index_ex = str(example_id)
+    example = example_data[index_ex]
+    paper_text = 'Title:\t' +  example['title'] + '\n\nAbstract:\t' + example['abstract']
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return [paper_text, tasks, gaps, keywords, recent_works]
+
+# WRONG: example_text = [load_example(1), load_example(2)]
+
+# ✅ FIXED:
+example_text = []
+for i in ["1", "2"]:
+    paper_text = 'Title:\t' + example_data[i]['title'] + '\n\nAbstract:\t' + example_data[i]['abstract']
+    example_text.append(paper_text)
+
+
+# Function to handle example clicks
+def load_example_and_set_index(paper_text_input):
+    global index_ex, state_example
+    state_example = True
+    index_ex = str(example_text.index(paper_text_input) + 1)
+    example =  load_example(index_ex)
+
+    return example
+
+
+
+########## Phase 1 ##############
+
+def extract_research_elements(paper_text):
+    global state_extract, index_ex, state_example
+    if not state_example or paper_text == "":
+        return "", "", "", ""
+    state_extract = True
+    if not paper_text.strip().startswith("Title:\t" + example_data[index_ex]["title"]):
+        print("Mismatch detected.")
+        print(paper_text)
+        return "", "", "", ""
+    example = example_data[index_ex]
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return tasks, gaps, keywords, recent_works
+
+
+# Step 2: Generate Research Hypothesis and Experiment Plan
+def generate_and_store(paper_text, tasks, gaps, keywords, recent_works):
+    if (not state_extract or not state_example or paper_text == ""):
+        return "", "", "", ""
+    global state_generate, index_ex
+    state_generate = True
+    hypothesis = example_data[index_ex]['hypothesis']
+    experiment_plan = example_data[index_ex]['experiment_plan']
+    return hypothesis, experiment_plan, hypothesis, experiment_plan
+
+##########  Phase 2 & 3  ##############
+def start_experiment_agent(hypothesis, plan):
+    if (not state_extract or not state_generate or not state_example):
+        return "", "", ""
+    global state_agent, step_index, state_complete
+    state_agent = True
+    step_index = 0
+    state_complete = False
+    # predefined_message = f"Implement the following hypothesis and experiment plan:\n\nHypothesis:\n{hypothesis}\n\nExperiment Plan:\n{plan}"
+    return example_data[index_ex]['code_init'], predefined_action_log, "", ""
+
+def submit_feedback(user_feedback, history, previous_response):
+    if (not state_extract or not  state_generate or not state_agent or not state_example):
+        return "", "", ""
+    global step_index, state_complete
+    step_index += 1
+    msg = history
+    if step_index < len(process_steps):
+        msg += previous_response + "\nUser feedback:" + user_feedback + "\n\n"
+        response_info = process_steps[step_index]
+        response = info_to_message(response_info)  # Convert dictionary to formatted string
+        response += "Please provide feedback based on the history, response entries, and observation, and questions: "
+        step_index += 1
+        msg += response
+    else:
+        state_complete = True
+        response = "Agent Finished."
+        
+    return msg, response, example_data[index_ex]['code_init'] if state_complete else example_data[index_ex]['code_final'], ""
+
+def load_phase_2_inputs(hypothesis, plan):
+    return hypothesis, plan, "# Code implementation will be displayed here after Start ExperimentAgent."
+
+
+
+predefined_action_log = """
+[Reasoning]: To understand the initial structure and functionality of train.py for effective improvements.
+[Action]: Inspect Script (train.py)
+Input: {"script_name": "train.py", "start_line_number": "1", "end_line_number": "74"}
+Objective: Understand the training script, including data processing, [...]
+[Observation]: The train.py script imports [...]. Sets random seeds [...]. Defines [...] Placeholder functions [...] exist without implementation. [...]
+[Feedback]: The script structure is clear, but key functions (train_model, predict) need proper implementation for proposed model training and prediction.\n
+"""
+
+
+predefined_observation = """
+Epoch [1/10],
+Train MSE: 0.543,
+Test MSE: 0.688
+Epoch [2/10],
+Train MSE: 0.242,
+Test MSE: 0.493\n
+"""
+
+# Initialize the global step_index and history
+process_steps = [
+    {
+        "Action": "Inspect Script Lines (train.py)",
+        "Observation": (
+            "The train.py script imports necessary libraries (e.g., pandas, sklearn, torch). "
+            "Sets random seeds for reproducibility. Defines compute_metrics_for_regression function "
+            "to calculate RMSE for different dimensions. Placeholder functions train_model and "
+            "predict exist without implementations."
+        ),
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The script executed successfully. Generated embeddings using the BERT model. Completed "
+            "the training process without errors. Metrics calculation placeholders indicated areas needing implementation."
+        ),
+    },
+    {
+        "Action": "Edit Script (train.py)",
+        "Observation": (
+            "Edited train.py to separate data loading, model definition, training loop, and evaluation into distinct functions. "
+            "The edited train.py now has clearly defined functions"
+            "for data loading (load_data), model definition (build_model), "
+            "training (train_model), and evaluation (evaluate_model). Similarly, eval.py is reorganized to load the model and perform predictions efficiently."
+        ),
+    },
+    {
+        "Action": "Retrieve Model",
+        "Observation": "CNN and BiLSTM retrieved.",
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The model trained over the specified number of epochs. Training and validation loss values are recorded for each epoch, "
+            "the decrease in loss indicates improved model performance."
+        )
+    },
+    {
+        "Action": "Evaluation",
+        "Observation": predefined_observation,
+    }
+]
+def info_to_message(info):
+    msg = ""
+    for k, v in info.items():
+        if isinstance(v, dict):
+            tempv = v
+            v = ""
+            for k2, v2 in tempv.items():
+                v += f"{k2}:\n  {v2}\n"
+        v = User.indent_text(v, 2)
+        msg += '-' * 64
+        msg += '\n'
+        msg += f"{k}:\n{v}\n"
+    return msg
+
+
+def handle_example_click(example_index):
+    global index_ex
+    index_ex = example_index
+    return load_example(index_ex)  # Simply return the text to display it in the textbox
+
+# Gradio Interface
+with gr.Blocks(css=".gr-examples-label {display: none;}", theme=gr.themes.Default()) as app:
+    gr.Markdown("# MLR- Copilot: Machine Learning Research based on LLM Agents")
+    gr.Markdown("###   ")
+    gr.Markdown("## <span style='color:Orange;'> This UI is for predefined example demo only.</span>")
+    gr.Markdown("## <span style='color:Orange;'> To reproduce the results please use [Github](https://github.com/du-nlp-lab/MLR-Copilot/)</span>")
+
+
+
+    gr.Markdown("MLR-Copilot is a framework where LLMs mimic researchers’ thought processes, designed to enhance the productivity of machine learning research by automating the generation and implementation of research ideas. It begins with a research paper, autonomously generating and validating these ideas, while incorporating human feedback to help reach executable research outcomes.")
+
+
+    gr.Markdown("##    ⬇️  Click an example at bottom to start！")
+
+
+    # Use state variables to store generated hypothesis and experiment plan
+    hypothesis_state = gr.State("")
+    experiment_plan_state = gr.State("")
+
+    ########## Phase 1:  Research Idea Generation Tab ##############
+    with gr.Tab("💡Stage 1: Research Idea Generation"):
+        gr.Markdown("### Extract Research Elements and Generate Research Ideas")
+        with gr.Row():
+            with gr.Column():
+                paper_text_input = gr.Textbox(value="", lines=10, label="📑 Research Paper Text", interactive=False)
+
+                # extract_button = gr.Button("🔍 Extract Research Elements")
+                with gr.Row():
+                    tasks_output = gr.Textbox(placeholder="Research task definition", label="Research Tasks", lines=2, interactive=False)
+                    gaps_output = gr.Textbox(placeholder="Research gaps of current works", label="Research Gaps", lines=2, interactive=False)
+                    keywords_output = gr.Textbox(placeholder="Paper keywords", label="Keywords", lines=2, interactive=False)
+                    recent_works_output = gr.Textbox(placeholder="Recent works extracted from Semantic Scholar", label="Recent Works", lines=2, interactive=False)
+            with gr.Column():
+                with gr.Row():  # Move the button to the top 
+                    generate_button = gr.Button("✍️ Generate Research Hypothesis & Experiment Plan")
+                with gr.Group():
+                    gr.Markdown("### 🌟 Research Idea")
+                    with gr.Row():
+                        hypothesis_output = gr.Textbox(label="Generated Hypothesis", lines=20, interactive=False)
+                        experiment_plan_output = gr.Textbox(label="Generated Experiment Plan", lines=20, interactive=False)
+        
+    
+        # Step 1: Extract Research Elements
+        # extract_button.click(
+        #     fn=extract_research_elements, 
+        #     inputs=paper_text_input, 
+        #     outputs=[tasks_output, gaps_output, keywords_output, recent_works_output]
+        # )
+
+        generate_button.click(
+            fn=generate_and_store, 
+            inputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output], 
+            outputs=[hypothesis_output, experiment_plan_output, hypothesis_state, experiment_plan_state]
+        )
+
+        gr.Examples(
+            examples=example_text,
+            inputs=[paper_text_input],
+            outputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output, hypothesis_output, experiment_plan_output],
+            fn=load_example_and_set_index,
+            run_on_click = True,
+            # label="⬇️ Click an example to load"
+        )
+        
+
+
+
+    ########## Phase 2 & 3:  Experiment implementation and execution ##############
+    with gr.Tab("🧪 Stage 2 & Stage 3: Experiment implementation and execution"):
+        gr.Markdown("### Interact with the ExperimentAgent")
+
+        with gr.Row():
+            with gr.Column():
+                with gr.Group():
+                    gr.Markdown("### 🌟 Generated Research Idea")
+                    with gr.Row():
+                        idea_input = gr.Textbox(label="Generated Research Hypothesis", lines=30, interactive=False)
+                        plan_input = gr.Textbox(label="Generated Experiment Plan", lines=30, interactive=False)
+        
+            with gr.Column():
+                start_exp_agnet = gr.Button("⚙️ Start / Reset ExperimentAgent", elem_classes=["agent-btn"])
+                with gr.Group():
+                    gr.Markdown("### Implementation + Execution Log")    
+                    log = gr.Textbox(label="📖 Execution Log", lines=20, interactive=False)
+                    code_display = gr.Code(label="🧑‍💻 Implementation", language="python", interactive=False)
+
+            with gr.Column():
+                response = gr.Textbox(label="🤖 ExperimentAgent Response", lines=30, interactive=False)
+                feedback = gr.Textbox(placeholder="N/A", label="🧑‍🔬 User Feedback", lines=3, interactive=True)
+                submit_button = gr.Button("Submit", elem_classes=["Submit-btn"])
+        
+        hypothesis_state.change(
+            fn=load_phase_2_inputs, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[idea_input, plan_input, code_display]
+        )
+        
+        # Start research agent
+        start_exp_agnet.click(
+            fn=start_experiment_agent, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[code_display, log, response, feedback]
+        )
+        
+        submit_button.click(
+            fn=submit_feedback,
+            inputs=[feedback, log, response],
+            outputs=[log, response, code_display, feedback]
+        )
+
+# Test
+if __name__ == "__main__":
+    step_index = 0
+    app.launch(share=True)

.history/app_20250404174352.py

@@ -0,0 +1,341 @@
+import gradio as gr
+from pathlib import Path
+from reactagent.environment import Environment
+from reactagent.agents.agent_research import ResearchAgent
+from reactagent.runner import create_parser
+from reactagent import llm
+from reactagent.users.user import User
+import os
+import json
+
+
+# Global variables to store session state
+env = None
+agent = None
+state_example = False
+state_extract = False
+state_generate = False
+state_agent = False
+state_complete = False
+index_ex = "1"
+
+example_text = [
+    "Research Paper 1: Dataset and Baseline for Automatic Student Feedback Analysis",
+    "Research Paper 2: An Empirical Study on the Impact of Code Review on Software Quality"
+]
+
+
+# Load example JSON file
+def load_example_data():
+    with open("example/example_data.json", "r") as json_file:
+        example_data = json.load(json_file)
+    
+    for idx in example_data.keys():
+        try:
+            file = example_data[idx]["code_init"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_init"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")
+        try:
+            file = example_data[idx]["code_final"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_final"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")     
+    return example_data
+
+example_data = load_example_data()
+
+# Function to handle the selection of an example and populate the respective fields
+def load_example(example_id):
+    global index_ex 
+    index_ex = str(example_id)
+    example = example_data[index_ex]
+    paper_text = 'Title:\t' +  example['title'] + '\n\nAbstract:\t' + example['abstract']
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return [paper_text, tasks, gaps, keywords, recent_works]
+
+    # WRONG: example_text = [load_example(1), load_example(2)]
+
+    # ✅ FIXED:
+    example_text = []
+    for i in ["1", "2"]:
+        paper_text = 'Title:\t' + example_data[i]['title'] + '\n\nAbstract:\t' + example_data[i]['abstract']
+        example_text.append(paper_text)
+
+
+# Function to handle example clicks
+def load_example_and_set_index(paper_text_input):
+    global index_ex, state_example
+    state_example = True
+    index_ex = str(example_text.index(paper_text_input) + 1)
+    example =  load_example(index_ex)
+
+    return example
+
+
+
+########## Phase 1 ##############
+
+def extract_research_elements(paper_text):
+    global state_extract, index_ex, state_example
+    if not state_example or paper_text == "":
+        return "", "", "", ""
+    state_extract = True
+    if not paper_text.strip().startswith("Title:\t" + example_data[index_ex]["title"]):
+        print("Mismatch detected.")
+        print(paper_text)
+        return "", "", "", ""
+    example = example_data[index_ex]
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return tasks, gaps, keywords, recent_works
+
+
+# Step 2: Generate Research Hypothesis and Experiment Plan
+def generate_and_store(paper_text, tasks, gaps, keywords, recent_works):
+    if (not state_extract or not state_example or paper_text == ""):
+        return "", "", "", ""
+    global state_generate, index_ex
+    state_generate = True
+    hypothesis = example_data[index_ex]['hypothesis']
+    experiment_plan = example_data[index_ex]['experiment_plan']
+    return hypothesis, experiment_plan, hypothesis, experiment_plan
+
+##########  Phase 2 & 3  ##############
+def start_experiment_agent(hypothesis, plan):
+    if (not state_extract or not state_generate or not state_example):
+        return "", "", ""
+    global state_agent, step_index, state_complete
+    state_agent = True
+    step_index = 0
+    state_complete = False
+    # predefined_message = f"Implement the following hypothesis and experiment plan:\n\nHypothesis:\n{hypothesis}\n\nExperiment Plan:\n{plan}"
+    return example_data[index_ex]['code_init'], predefined_action_log, "", ""
+
+def submit_feedback(user_feedback, history, previous_response):
+    if (not state_extract or not  state_generate or not state_agent or not state_example):
+        return "", "", ""
+    global step_index, state_complete
+    step_index += 1
+    msg = history
+    if step_index < len(process_steps):
+        msg += previous_response + "\nUser feedback:" + user_feedback + "\n\n"
+        response_info = process_steps[step_index]
+        response = info_to_message(response_info)  # Convert dictionary to formatted string
+        response += "Please provide feedback based on the history, response entries, and observation, and questions: "
+        step_index += 1
+        msg += response
+    else:
+        state_complete = True
+        response = "Agent Finished."
+        
+    return msg, response, example_data[index_ex]['code_init'] if state_complete else example_data[index_ex]['code_final'], ""
+
+def load_phase_2_inputs(hypothesis, plan):
+    return hypothesis, plan, "# Code implementation will be displayed here after Start ExperimentAgent."
+
+
+
+predefined_action_log = """
+[Reasoning]: To understand the initial structure and functionality of train.py for effective improvements.
+[Action]: Inspect Script (train.py)
+Input: {"script_name": "train.py", "start_line_number": "1", "end_line_number": "74"}
+Objective: Understand the training script, including data processing, [...]
+[Observation]: The train.py script imports [...]. Sets random seeds [...]. Defines [...] Placeholder functions [...] exist without implementation. [...]
+[Feedback]: The script structure is clear, but key functions (train_model, predict) need proper implementation for proposed model training and prediction.\n
+"""
+
+
+predefined_observation = """
+Epoch [1/10],
+Train MSE: 0.543,
+Test MSE: 0.688
+Epoch [2/10],
+Train MSE: 0.242,
+Test MSE: 0.493\n
+"""
+
+# Initialize the global step_index and history
+process_steps = [
+    {
+        "Action": "Inspect Script Lines (train.py)",
+        "Observation": (
+            "The train.py script imports necessary libraries (e.g., pandas, sklearn, torch). "
+            "Sets random seeds for reproducibility. Defines compute_metrics_for_regression function "
+            "to calculate RMSE for different dimensions. Placeholder functions train_model and "
+            "predict exist without implementations."
+        ),
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The script executed successfully. Generated embeddings using the BERT model. Completed "
+            "the training process without errors. Metrics calculation placeholders indicated areas needing implementation."
+        ),
+    },
+    {
+        "Action": "Edit Script (train.py)",
+        "Observation": (
+            "Edited train.py to separate data loading, model definition, training loop, and evaluation into distinct functions. "
+            "The edited train.py now has clearly defined functions"
+            "for data loading (load_data), model definition (build_model), "
+            "training (train_model), and evaluation (evaluate_model). Similarly, eval.py is reorganized to load the model and perform predictions efficiently."
+        ),
+    },
+    {
+        "Action": "Retrieve Model",
+        "Observation": "CNN and BiLSTM retrieved.",
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The model trained over the specified number of epochs. Training and validation loss values are recorded for each epoch, "
+            "the decrease in loss indicates improved model performance."
+        )
+    },
+    {
+        "Action": "Evaluation",
+        "Observation": predefined_observation,
+    }
+]
+def info_to_message(info):
+    msg = ""
+    for k, v in info.items():
+        if isinstance(v, dict):
+            tempv = v
+            v = ""
+            for k2, v2 in tempv.items():
+                v += f"{k2}:\n  {v2}\n"
+        v = User.indent_text(v, 2)
+        msg += '-' * 64
+        msg += '\n'
+        msg += f"{k}:\n{v}\n"
+    return msg
+
+
+def handle_example_click(example_index):
+    global index_ex
+    index_ex = example_index
+    return load_example(index_ex)  # Simply return the text to display it in the textbox
+
+# Gradio Interface
+with gr.Blocks(css=".gr-examples-label {display: none;}", theme=gr.themes.Default()) as app:
+    gr.Markdown("# MLR- Copilot: Machine Learning Research based on LLM Agents")
+    gr.Markdown("###   ")
+    gr.Markdown("## <span style='color:Orange;'> This UI is for predefined example demo only.</span>")
+    gr.Markdown("## <span style='color:Orange;'> To reproduce the results please use [Github](https://github.com/du-nlp-lab/MLR-Copilot/)</span>")
+
+
+
+    gr.Markdown("MLR-Copilot is a framework where LLMs mimic researchers’ thought processes, designed to enhance the productivity of machine learning research by automating the generation and implementation of research ideas. It begins with a research paper, autonomously generating and validating these ideas, while incorporating human feedback to help reach executable research outcomes.")
+
+
+    gr.Markdown("##    ⬇️  Click an example at bottom to start！")
+
+
+    # Use state variables to store generated hypothesis and experiment plan
+    hypothesis_state = gr.State("")
+    experiment_plan_state = gr.State("")
+
+    ########## Phase 1:  Research Idea Generation Tab ##############
+    with gr.Tab("💡Stage 1: Research Idea Generation"):
+        gr.Markdown("### Extract Research Elements and Generate Research Ideas")
+        with gr.Row():
+            with gr.Column():
+                paper_text_input = gr.Textbox(value="", lines=10, label="📑 Research Paper Text", interactive=False)
+
+                # extract_button = gr.Button("🔍 Extract Research Elements")
+                with gr.Row():
+                    tasks_output = gr.Textbox(placeholder="Research task definition", label="Research Tasks", lines=2, interactive=False)
+                    gaps_output = gr.Textbox(placeholder="Research gaps of current works", label="Research Gaps", lines=2, interactive=False)
+                    keywords_output = gr.Textbox(placeholder="Paper keywords", label="Keywords", lines=2, interactive=False)
+                    recent_works_output = gr.Textbox(placeholder="Recent works extracted from Semantic Scholar", label="Recent Works", lines=2, interactive=False)
+            with gr.Column():
+                with gr.Row():  # Move the button to the top 
+                    generate_button = gr.Button("✍️ Generate Research Hypothesis & Experiment Plan")
+                with gr.Group():
+                    gr.Markdown("### 🌟 Research Idea")
+                    with gr.Row():
+                        hypothesis_output = gr.Textbox(label="Generated Hypothesis", lines=20, interactive=False)
+                        experiment_plan_output = gr.Textbox(label="Generated Experiment Plan", lines=20, interactive=False)
+        
+    
+        # Step 1: Extract Research Elements
+        # extract_button.click(
+        #     fn=extract_research_elements, 
+        #     inputs=paper_text_input, 
+        #     outputs=[tasks_output, gaps_output, keywords_output, recent_works_output]
+        # )
+
+        generate_button.click(
+            fn=generate_and_store, 
+            inputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output], 
+            outputs=[hypothesis_output, experiment_plan_output, hypothesis_state, experiment_plan_state]
+        )
+
+        gr.Examples(
+            examples=example_text,
+            inputs=[paper_text_input],
+            outputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output, hypothesis_output, experiment_plan_output],
+            fn=load_example_and_set_index,
+            run_on_click = True,
+            # label="⬇️ Click an example to load"
+        )
+        
+
+
+
+    ########## Phase 2 & 3:  Experiment implementation and execution ##############
+    with gr.Tab("🧪 Stage 2 & Stage 3: Experiment implementation and execution"):
+        gr.Markdown("### Interact with the ExperimentAgent")
+
+        with gr.Row():
+            with gr.Column():
+                with gr.Group():
+                    gr.Markdown("### 🌟 Generated Research Idea")
+                    with gr.Row():
+                        idea_input = gr.Textbox(label="Generated Research Hypothesis", lines=30, interactive=False)
+                        plan_input = gr.Textbox(label="Generated Experiment Plan", lines=30, interactive=False)
+        
+            with gr.Column():
+                start_exp_agnet = gr.Button("⚙️ Start / Reset ExperimentAgent", elem_classes=["agent-btn"])
+                with gr.Group():
+                    gr.Markdown("### Implementation + Execution Log")    
+                    log = gr.Textbox(label="📖 Execution Log", lines=20, interactive=False)
+                    code_display = gr.Code(label="🧑‍💻 Implementation", language="python", interactive=False)
+
+            with gr.Column():
+                response = gr.Textbox(label="🤖 ExperimentAgent Response", lines=30, interactive=False)
+                feedback = gr.Textbox(placeholder="N/A", label="🧑‍🔬 User Feedback", lines=3, interactive=True)
+                submit_button = gr.Button("Submit", elem_classes=["Submit-btn"])
+        
+        hypothesis_state.change(
+            fn=load_phase_2_inputs, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[idea_input, plan_input, code_display]
+        )
+        
+        # Start research agent
+        start_exp_agnet.click(
+            fn=start_experiment_agent, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[code_display, log, response, feedback]
+        )
+        
+        submit_button.click(
+            fn=submit_feedback,
+            inputs=[feedback, log, response],
+            outputs=[log, response, code_display, feedback]
+        )
+
+# Test
+if __name__ == "__main__":
+    step_index = 0
+    app.launch(share=True)

.history/app_20250404174355.py

@@ -0,0 +1,340 @@
+import gradio as gr
+from pathlib import Path
+from reactagent.environment import Environment
+from reactagent.agents.agent_research import ResearchAgent
+from reactagent.runner import create_parser
+from reactagent import llm
+from reactagent.users.user import User
+import os
+import json
+
+
+# Global variables to store session state
+env = None
+agent = None
+state_example = False
+state_extract = False
+state_generate = False
+state_agent = False
+state_complete = False
+index_ex = "1"
+
+example_text = [
+    "Research Paper 1: Dataset and Baseline for Automatic Student Feedback Analysis",
+    "Research Paper 2: An Empirical Study on the Impact of Code Review on Software Quality"
+]
+
+
+# Load example JSON file
+def load_example_data():
+    with open("example/example_data.json", "r") as json_file:
+        example_data = json.load(json_file)
+    
+    for idx in example_data.keys():
+        try:
+            file = example_data[idx]["code_init"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_init"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")
+        try:
+            file = example_data[idx]["code_final"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_final"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")     
+    return example_data
+
+example_data = load_example_data()
+
+# Function to handle the selection of an example and populate the respective fields
+def load_example(example_id):
+    global index_ex 
+    index_ex = str(example_id)
+    example = example_data[index_ex]
+    paper_text = 'Title:\t' +  example['title'] + '\n\nAbstract:\t' + example['abstract']
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return [paper_text, tasks, gaps, keywords, recent_works]
+
+    # WRONG: example_text = [load_example(1), load_example(2)]
+    # ✅ FIXED:
+    example_text = []
+    for i in ["1", "2"]:
+        paper_text = 'Title:\t' + example_data[i]['title'] + '\n\nAbstract:\t' + example_data[i]['abstract']
+        example_text.append(paper_text)
+
+
+# Function to handle example clicks
+def load_example_and_set_index(paper_text_input):
+    global index_ex, state_example
+    state_example = True
+    index_ex = str(example_text.index(paper_text_input) + 1)
+    example =  load_example(index_ex)
+
+    return example
+
+
+
+########## Phase 1 ##############
+
+def extract_research_elements(paper_text):
+    global state_extract, index_ex, state_example
+    if not state_example or paper_text == "":
+        return "", "", "", ""
+    state_extract = True
+    if not paper_text.strip().startswith("Title:\t" + example_data[index_ex]["title"]):
+        print("Mismatch detected.")
+        print(paper_text)
+        return "", "", "", ""
+    example = example_data[index_ex]
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return tasks, gaps, keywords, recent_works
+
+
+# Step 2: Generate Research Hypothesis and Experiment Plan
+def generate_and_store(paper_text, tasks, gaps, keywords, recent_works):
+    if (not state_extract or not state_example or paper_text == ""):
+        return "", "", "", ""
+    global state_generate, index_ex
+    state_generate = True
+    hypothesis = example_data[index_ex]['hypothesis']
+    experiment_plan = example_data[index_ex]['experiment_plan']
+    return hypothesis, experiment_plan, hypothesis, experiment_plan
+
+##########  Phase 2 & 3  ##############
+def start_experiment_agent(hypothesis, plan):
+    if (not state_extract or not state_generate or not state_example):
+        return "", "", ""
+    global state_agent, step_index, state_complete
+    state_agent = True
+    step_index = 0
+    state_complete = False
+    # predefined_message = f"Implement the following hypothesis and experiment plan:\n\nHypothesis:\n{hypothesis}\n\nExperiment Plan:\n{plan}"
+    return example_data[index_ex]['code_init'], predefined_action_log, "", ""
+
+def submit_feedback(user_feedback, history, previous_response):
+    if (not state_extract or not  state_generate or not state_agent or not state_example):
+        return "", "", ""
+    global step_index, state_complete
+    step_index += 1
+    msg = history
+    if step_index < len(process_steps):
+        msg += previous_response + "\nUser feedback:" + user_feedback + "\n\n"
+        response_info = process_steps[step_index]
+        response = info_to_message(response_info)  # Convert dictionary to formatted string
+        response += "Please provide feedback based on the history, response entries, and observation, and questions: "
+        step_index += 1
+        msg += response
+    else:
+        state_complete = True
+        response = "Agent Finished."
+        
+    return msg, response, example_data[index_ex]['code_init'] if state_complete else example_data[index_ex]['code_final'], ""
+
+def load_phase_2_inputs(hypothesis, plan):
+    return hypothesis, plan, "# Code implementation will be displayed here after Start ExperimentAgent."
+
+
+
+predefined_action_log = """
+[Reasoning]: To understand the initial structure and functionality of train.py for effective improvements.
+[Action]: Inspect Script (train.py)
+Input: {"script_name": "train.py", "start_line_number": "1", "end_line_number": "74"}
+Objective: Understand the training script, including data processing, [...]
+[Observation]: The train.py script imports [...]. Sets random seeds [...]. Defines [...] Placeholder functions [...] exist without implementation. [...]
+[Feedback]: The script structure is clear, but key functions (train_model, predict) need proper implementation for proposed model training and prediction.\n
+"""
+
+
+predefined_observation = """
+Epoch [1/10],
+Train MSE: 0.543,
+Test MSE: 0.688
+Epoch [2/10],
+Train MSE: 0.242,
+Test MSE: 0.493\n
+"""
+
+# Initialize the global step_index and history
+process_steps = [
+    {
+        "Action": "Inspect Script Lines (train.py)",
+        "Observation": (
+            "The train.py script imports necessary libraries (e.g., pandas, sklearn, torch). "
+            "Sets random seeds for reproducibility. Defines compute_metrics_for_regression function "
+            "to calculate RMSE for different dimensions. Placeholder functions train_model and "
+            "predict exist without implementations."
+        ),
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The script executed successfully. Generated embeddings using the BERT model. Completed "
+            "the training process without errors. Metrics calculation placeholders indicated areas needing implementation."
+        ),
+    },
+    {
+        "Action": "Edit Script (train.py)",
+        "Observation": (
+            "Edited train.py to separate data loading, model definition, training loop, and evaluation into distinct functions. "
+            "The edited train.py now has clearly defined functions"
+            "for data loading (load_data), model definition (build_model), "
+            "training (train_model), and evaluation (evaluate_model). Similarly, eval.py is reorganized to load the model and perform predictions efficiently."
+        ),
+    },
+    {
+        "Action": "Retrieve Model",
+        "Observation": "CNN and BiLSTM retrieved.",
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The model trained over the specified number of epochs. Training and validation loss values are recorded for each epoch, "
+            "the decrease in loss indicates improved model performance."
+        )
+    },
+    {
+        "Action": "Evaluation",
+        "Observation": predefined_observation,
+    }
+]
+def info_to_message(info):
+    msg = ""
+    for k, v in info.items():
+        if isinstance(v, dict):
+            tempv = v
+            v = ""
+            for k2, v2 in tempv.items():
+                v += f"{k2}:\n  {v2}\n"
+        v = User.indent_text(v, 2)
+        msg += '-' * 64
+        msg += '\n'
+        msg += f"{k}:\n{v}\n"
+    return msg
+
+
+def handle_example_click(example_index):
+    global index_ex
+    index_ex = example_index
+    return load_example(index_ex)  # Simply return the text to display it in the textbox
+
+# Gradio Interface
+with gr.Blocks(css=".gr-examples-label {display: none;}", theme=gr.themes.Default()) as app:
+    gr.Markdown("# MLR- Copilot: Machine Learning Research based on LLM Agents")
+    gr.Markdown("###   ")
+    gr.Markdown("## <span style='color:Orange;'> This UI is for predefined example demo only.</span>")
+    gr.Markdown("## <span style='color:Orange;'> To reproduce the results please use [Github](https://github.com/du-nlp-lab/MLR-Copilot/)</span>")
+
+
+
+    gr.Markdown("MLR-Copilot is a framework where LLMs mimic researchers’ thought processes, designed to enhance the productivity of machine learning research by automating the generation and implementation of research ideas. It begins with a research paper, autonomously generating and validating these ideas, while incorporating human feedback to help reach executable research outcomes.")
+
+
+    gr.Markdown("##    ⬇️  Click an example at bottom to start！")
+
+
+    # Use state variables to store generated hypothesis and experiment plan
+    hypothesis_state = gr.State("")
+    experiment_plan_state = gr.State("")
+
+    ########## Phase 1:  Research Idea Generation Tab ##############
+    with gr.Tab("💡Stage 1: Research Idea Generation"):
+        gr.Markdown("### Extract Research Elements and Generate Research Ideas")
+        with gr.Row():
+            with gr.Column():
+                paper_text_input = gr.Textbox(value="", lines=10, label="📑 Research Paper Text", interactive=False)
+
+                # extract_button = gr.Button("🔍 Extract Research Elements")
+                with gr.Row():
+                    tasks_output = gr.Textbox(placeholder="Research task definition", label="Research Tasks", lines=2, interactive=False)
+                    gaps_output = gr.Textbox(placeholder="Research gaps of current works", label="Research Gaps", lines=2, interactive=False)
+                    keywords_output = gr.Textbox(placeholder="Paper keywords", label="Keywords", lines=2, interactive=False)
+                    recent_works_output = gr.Textbox(placeholder="Recent works extracted from Semantic Scholar", label="Recent Works", lines=2, interactive=False)
+            with gr.Column():
+                with gr.Row():  # Move the button to the top 
+                    generate_button = gr.Button("✍️ Generate Research Hypothesis & Experiment Plan")
+                with gr.Group():
+                    gr.Markdown("### 🌟 Research Idea")
+                    with gr.Row():
+                        hypothesis_output = gr.Textbox(label="Generated Hypothesis", lines=20, interactive=False)
+                        experiment_plan_output = gr.Textbox(label="Generated Experiment Plan", lines=20, interactive=False)
+        
+    
+        # Step 1: Extract Research Elements
+        # extract_button.click(
+        #     fn=extract_research_elements, 
+        #     inputs=paper_text_input, 
+        #     outputs=[tasks_output, gaps_output, keywords_output, recent_works_output]
+        # )
+
+        generate_button.click(
+            fn=generate_and_store, 
+            inputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output], 
+            outputs=[hypothesis_output, experiment_plan_output, hypothesis_state, experiment_plan_state]
+        )
+
+        gr.Examples(
+            examples=example_text,
+            inputs=[paper_text_input],
+            outputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output, hypothesis_output, experiment_plan_output],
+            fn=load_example_and_set_index,
+            run_on_click = True,
+            # label="⬇️ Click an example to load"
+        )
+        
+
+
+
+    ########## Phase 2 & 3:  Experiment implementation and execution ##############
+    with gr.Tab("🧪 Stage 2 & Stage 3: Experiment implementation and execution"):
+        gr.Markdown("### Interact with the ExperimentAgent")
+
+        with gr.Row():
+            with gr.Column():
+                with gr.Group():
+                    gr.Markdown("### 🌟 Generated Research Idea")
+                    with gr.Row():
+                        idea_input = gr.Textbox(label="Generated Research Hypothesis", lines=30, interactive=False)
+                        plan_input = gr.Textbox(label="Generated Experiment Plan", lines=30, interactive=False)
+        
+            with gr.Column():
+                start_exp_agnet = gr.Button("⚙️ Start / Reset ExperimentAgent", elem_classes=["agent-btn"])
+                with gr.Group():
+                    gr.Markdown("### Implementation + Execution Log")    
+                    log = gr.Textbox(label="📖 Execution Log", lines=20, interactive=False)
+                    code_display = gr.Code(label="🧑‍💻 Implementation", language="python", interactive=False)
+
+            with gr.Column():
+                response = gr.Textbox(label="🤖 ExperimentAgent Response", lines=30, interactive=False)
+                feedback = gr.Textbox(placeholder="N/A", label="🧑‍🔬 User Feedback", lines=3, interactive=True)
+                submit_button = gr.Button("Submit", elem_classes=["Submit-btn"])
+        
+        hypothesis_state.change(
+            fn=load_phase_2_inputs, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[idea_input, plan_input, code_display]
+        )
+        
+        # Start research agent
+        start_exp_agnet.click(
+            fn=start_experiment_agent, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[code_display, log, response, feedback]
+        )
+        
+        submit_button.click(
+            fn=submit_feedback,
+            inputs=[feedback, log, response],
+            outputs=[log, response, code_display, feedback]
+        )
+
+# Test
+if __name__ == "__main__":
+    step_index = 0
+    app.launch(share=True)

.history/app_20250404174357.py

@@ -0,0 +1,339 @@
+import gradio as gr
+from pathlib import Path
+from reactagent.environment import Environment
+from reactagent.agents.agent_research import ResearchAgent
+from reactagent.runner import create_parser
+from reactagent import llm
+from reactagent.users.user import User
+import os
+import json
+
+
+# Global variables to store session state
+env = None
+agent = None
+state_example = False
+state_extract = False
+state_generate = False
+state_agent = False
+state_complete = False
+index_ex = "1"
+
+example_text = [
+    "Research Paper 1: Dataset and Baseline for Automatic Student Feedback Analysis",
+    "Research Paper 2: An Empirical Study on the Impact of Code Review on Software Quality"
+]
+
+
+# Load example JSON file
+def load_example_data():
+    with open("example/example_data.json", "r") as json_file:
+        example_data = json.load(json_file)
+    
+    for idx in example_data.keys():
+        try:
+            file = example_data[idx]["code_init"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_init"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")
+        try:
+            file = example_data[idx]["code_final"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_final"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")     
+    return example_data
+
+example_data = load_example_data()
+
+# Function to handle the selection of an example and populate the respective fields
+def load_example(example_id):
+    global index_ex 
+    index_ex = str(example_id)
+    example = example_data[index_ex]
+    paper_text = 'Title:\t' +  example['title'] + '\n\nAbstract:\t' + example['abstract']
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return [paper_text, tasks, gaps, keywords, recent_works]
+
+    # WRONG: example_text = [load_example(1), load_example(2)]
+    example_text = []
+    for i in ["1", "2"]:
+        paper_text = 'Title:\t' + example_data[i]['title'] + '\n\nAbstract:\t' + example_data[i]['abstract']
+        example_text.append(paper_text)
+
+
+# Function to handle example clicks
+def load_example_and_set_index(paper_text_input):
+    global index_ex, state_example
+    state_example = True
+    index_ex = str(example_text.index(paper_text_input) + 1)
+    example =  load_example(index_ex)
+
+    return example
+
+
+
+########## Phase 1 ##############
+
+def extract_research_elements(paper_text):
+    global state_extract, index_ex, state_example
+    if not state_example or paper_text == "":
+        return "", "", "", ""
+    state_extract = True
+    if not paper_text.strip().startswith("Title:\t" + example_data[index_ex]["title"]):
+        print("Mismatch detected.")
+        print(paper_text)
+        return "", "", "", ""
+    example = example_data[index_ex]
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return tasks, gaps, keywords, recent_works
+
+
+# Step 2: Generate Research Hypothesis and Experiment Plan
+def generate_and_store(paper_text, tasks, gaps, keywords, recent_works):
+    if (not state_extract or not state_example or paper_text == ""):
+        return "", "", "", ""
+    global state_generate, index_ex
+    state_generate = True
+    hypothesis = example_data[index_ex]['hypothesis']
+    experiment_plan = example_data[index_ex]['experiment_plan']
+    return hypothesis, experiment_plan, hypothesis, experiment_plan
+
+##########  Phase 2 & 3  ##############
+def start_experiment_agent(hypothesis, plan):
+    if (not state_extract or not state_generate or not state_example):
+        return "", "", ""
+    global state_agent, step_index, state_complete
+    state_agent = True
+    step_index = 0
+    state_complete = False
+    # predefined_message = f"Implement the following hypothesis and experiment plan:\n\nHypothesis:\n{hypothesis}\n\nExperiment Plan:\n{plan}"
+    return example_data[index_ex]['code_init'], predefined_action_log, "", ""
+
+def submit_feedback(user_feedback, history, previous_response):
+    if (not state_extract or not  state_generate or not state_agent or not state_example):
+        return "", "", ""
+    global step_index, state_complete
+    step_index += 1
+    msg = history
+    if step_index < len(process_steps):
+        msg += previous_response + "\nUser feedback:" + user_feedback + "\n\n"
+        response_info = process_steps[step_index]
+        response = info_to_message(response_info)  # Convert dictionary to formatted string
+        response += "Please provide feedback based on the history, response entries, and observation, and questions: "
+        step_index += 1
+        msg += response
+    else:
+        state_complete = True
+        response = "Agent Finished."
+        
+    return msg, response, example_data[index_ex]['code_init'] if state_complete else example_data[index_ex]['code_final'], ""
+
+def load_phase_2_inputs(hypothesis, plan):
+    return hypothesis, plan, "# Code implementation will be displayed here after Start ExperimentAgent."
+
+
+
+predefined_action_log = """
+[Reasoning]: To understand the initial structure and functionality of train.py for effective improvements.
+[Action]: Inspect Script (train.py)
+Input: {"script_name": "train.py", "start_line_number": "1", "end_line_number": "74"}
+Objective: Understand the training script, including data processing, [...]
+[Observation]: The train.py script imports [...]. Sets random seeds [...]. Defines [...] Placeholder functions [...] exist without implementation. [...]
+[Feedback]: The script structure is clear, but key functions (train_model, predict) need proper implementation for proposed model training and prediction.\n
+"""
+
+
+predefined_observation = """
+Epoch [1/10],
+Train MSE: 0.543,
+Test MSE: 0.688
+Epoch [2/10],
+Train MSE: 0.242,
+Test MSE: 0.493\n
+"""
+
+# Initialize the global step_index and history
+process_steps = [
+    {
+        "Action": "Inspect Script Lines (train.py)",
+        "Observation": (
+            "The train.py script imports necessary libraries (e.g., pandas, sklearn, torch). "
+            "Sets random seeds for reproducibility. Defines compute_metrics_for_regression function "
+            "to calculate RMSE for different dimensions. Placeholder functions train_model and "
+            "predict exist without implementations."
+        ),
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The script executed successfully. Generated embeddings using the BERT model. Completed "
+            "the training process without errors. Metrics calculation placeholders indicated areas needing implementation."
+        ),
+    },
+    {
+        "Action": "Edit Script (train.py)",
+        "Observation": (
+            "Edited train.py to separate data loading, model definition, training loop, and evaluation into distinct functions. "
+            "The edited train.py now has clearly defined functions"
+            "for data loading (load_data), model definition (build_model), "
+            "training (train_model), and evaluation (evaluate_model). Similarly, eval.py is reorganized to load the model and perform predictions efficiently."
+        ),
+    },
+    {
+        "Action": "Retrieve Model",
+        "Observation": "CNN and BiLSTM retrieved.",
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The model trained over the specified number of epochs. Training and validation loss values are recorded for each epoch, "
+            "the decrease in loss indicates improved model performance."
+        )
+    },
+    {
+        "Action": "Evaluation",
+        "Observation": predefined_observation,
+    }
+]
+def info_to_message(info):
+    msg = ""
+    for k, v in info.items():
+        if isinstance(v, dict):
+            tempv = v
+            v = ""
+            for k2, v2 in tempv.items():
+                v += f"{k2}:\n  {v2}\n"
+        v = User.indent_text(v, 2)
+        msg += '-' * 64
+        msg += '\n'
+        msg += f"{k}:\n{v}\n"
+    return msg
+
+
+def handle_example_click(example_index):
+    global index_ex
+    index_ex = example_index
+    return load_example(index_ex)  # Simply return the text to display it in the textbox
+
+# Gradio Interface
+with gr.Blocks(css=".gr-examples-label {display: none;}", theme=gr.themes.Default()) as app:
+    gr.Markdown("# MLR- Copilot: Machine Learning Research based on LLM Agents")
+    gr.Markdown("###   ")
+    gr.Markdown("## <span style='color:Orange;'> This UI is for predefined example demo only.</span>")
+    gr.Markdown("## <span style='color:Orange;'> To reproduce the results please use [Github](https://github.com/du-nlp-lab/MLR-Copilot/)</span>")
+
+
+
+    gr.Markdown("MLR-Copilot is a framework where LLMs mimic researchers’ thought processes, designed to enhance the productivity of machine learning research by automating the generation and implementation of research ideas. It begins with a research paper, autonomously generating and validating these ideas, while incorporating human feedback to help reach executable research outcomes.")
+
+
+    gr.Markdown("##    ⬇️  Click an example at bottom to start！")
+
+
+    # Use state variables to store generated hypothesis and experiment plan
+    hypothesis_state = gr.State("")
+    experiment_plan_state = gr.State("")
+
+    ########## Phase 1:  Research Idea Generation Tab ##############
+    with gr.Tab("💡Stage 1: Research Idea Generation"):
+        gr.Markdown("### Extract Research Elements and Generate Research Ideas")
+        with gr.Row():
+            with gr.Column():
+                paper_text_input = gr.Textbox(value="", lines=10, label="📑 Research Paper Text", interactive=False)
+
+                # extract_button = gr.Button("🔍 Extract Research Elements")
+                with gr.Row():
+                    tasks_output = gr.Textbox(placeholder="Research task definition", label="Research Tasks", lines=2, interactive=False)
+                    gaps_output = gr.Textbox(placeholder="Research gaps of current works", label="Research Gaps", lines=2, interactive=False)
+                    keywords_output = gr.Textbox(placeholder="Paper keywords", label="Keywords", lines=2, interactive=False)
+                    recent_works_output = gr.Textbox(placeholder="Recent works extracted from Semantic Scholar", label="Recent Works", lines=2, interactive=False)
+            with gr.Column():
+                with gr.Row():  # Move the button to the top 
+                    generate_button = gr.Button("✍️ Generate Research Hypothesis & Experiment Plan")
+                with gr.Group():
+                    gr.Markdown("### 🌟 Research Idea")
+                    with gr.Row():
+                        hypothesis_output = gr.Textbox(label="Generated Hypothesis", lines=20, interactive=False)
+                        experiment_plan_output = gr.Textbox(label="Generated Experiment Plan", lines=20, interactive=False)
+        
+    
+        # Step 1: Extract Research Elements
+        # extract_button.click(
+        #     fn=extract_research_elements, 
+        #     inputs=paper_text_input, 
+        #     outputs=[tasks_output, gaps_output, keywords_output, recent_works_output]
+        # )
+
+        generate_button.click(
+            fn=generate_and_store, 
+            inputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output], 
+            outputs=[hypothesis_output, experiment_plan_output, hypothesis_state, experiment_plan_state]
+        )
+
+        gr.Examples(
+            examples=example_text,
+            inputs=[paper_text_input],
+            outputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output, hypothesis_output, experiment_plan_output],
+            fn=load_example_and_set_index,
+            run_on_click = True,
+            # label="⬇️ Click an example to load"
+        )
+        
+
+
+
+    ########## Phase 2 & 3:  Experiment implementation and execution ##############
+    with gr.Tab("🧪 Stage 2 & Stage 3: Experiment implementation and execution"):
+        gr.Markdown("### Interact with the ExperimentAgent")
+
+        with gr.Row():
+            with gr.Column():
+                with gr.Group():
+                    gr.Markdown("### 🌟 Generated Research Idea")
+                    with gr.Row():
+                        idea_input = gr.Textbox(label="Generated Research Hypothesis", lines=30, interactive=False)
+                        plan_input = gr.Textbox(label="Generated Experiment Plan", lines=30, interactive=False)
+        
+            with gr.Column():
+                start_exp_agnet = gr.Button("⚙️ Start / Reset ExperimentAgent", elem_classes=["agent-btn"])
+                with gr.Group():
+                    gr.Markdown("### Implementation + Execution Log")    
+                    log = gr.Textbox(label="📖 Execution Log", lines=20, interactive=False)
+                    code_display = gr.Code(label="🧑‍💻 Implementation", language="python", interactive=False)
+
+            with gr.Column():
+                response = gr.Textbox(label="🤖 ExperimentAgent Response", lines=30, interactive=False)
+                feedback = gr.Textbox(placeholder="N/A", label="🧑‍🔬 User Feedback", lines=3, interactive=True)
+                submit_button = gr.Button("Submit", elem_classes=["Submit-btn"])
+        
+        hypothesis_state.change(
+            fn=load_phase_2_inputs, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[idea_input, plan_input, code_display]
+        )
+        
+        # Start research agent
+        start_exp_agnet.click(
+            fn=start_experiment_agent, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[code_display, log, response, feedback]
+        )
+        
+        submit_button.click(
+            fn=submit_feedback,
+            inputs=[feedback, log, response],
+            outputs=[log, response, code_display, feedback]
+        )
+
+# Test
+if __name__ == "__main__":
+    step_index = 0
+    app.launch(share=True)

.history/app_20250404174402.py

@@ -0,0 +1,339 @@
+import gradio as gr
+from pathlib import Path
+from reactagent.environment import Environment
+from reactagent.agents.agent_research import ResearchAgent
+from reactagent.runner import create_parser
+from reactagent import llm
+from reactagent.users.user import User
+import os
+import json
+
+
+# Global variables to store session state
+env = None
+agent = None
+state_example = False
+state_extract = False
+state_generate = False
+state_agent = False
+state_complete = False
+index_ex = "1"
+
+example_text = [
+    "Research Paper 1: Dataset and Baseline for Automatic Student Feedback Analysis",
+    "Research Paper 2: An Empirical Study on the Impact of Code Review on Software Quality"
+]
+
+
+# Load example JSON file
+def load_example_data():
+    with open("example/example_data.json", "r") as json_file:
+        example_data = json.load(json_file)
+    
+    for idx in example_data.keys():
+        try:
+            file = example_data[idx]["code_init"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_init"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")
+        try:
+            file = example_data[idx]["code_final"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_final"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")     
+    return example_data
+
+example_data = load_example_data()
+
+# Function to handle the selection of an example and populate the respective fields
+def load_example(example_id):
+    global index_ex 
+    index_ex = str(example_id)
+    example = example_data[index_ex]
+    paper_text = 'Title:\t' +  example['title'] + '\n\nAbstract:\t' + example['abstract']
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return [paper_text, tasks, gaps, keywords, recent_works]
+
+    # WRONG: example_text = [load_example(1), load_example(2)]
+    example_text = []
+    for i in ["1", "2"]:
+        paper_text = 'Title:\t' + example_data[i]['title'] + '\n\nAbstract:\t' + example_data[i]['abstract']
+        example_text.append(paper_text)
+
+
+# Function to handle example clicks
+def load_example_and_set_index(paper_text_input):
+    global index_ex, state_example
+    state_example = True
+    index_ex = str(example_text.index(paper_text_input) + 1)
+    example =  load_example(index_ex)
+
+    return example
+
+
+
+########## Phase 1 ##############
+
+def extract_research_elements(paper_text):
+    global state_extract, index_ex, state_example
+    if not state_example or paper_text == "":
+        return "", "", "", ""
+    state_extract = True
+    if not paper_text.strip().startswith("Title:\t" + example_data[index_ex]["title"]):
+        print("Mismatch detected.")
+        print(paper_text)
+        return "", "", "", ""
+    example = example_data[index_ex]
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return tasks, gaps, keywords, recent_works
+
+
+# Step 2: Generate Research Hypothesis and Experiment Plan
+def generate_and_store(paper_text, tasks, gaps, keywords, recent_works):
+    if (not state_extract or not state_example or paper_text == ""):
+        return "", "", "", ""
+    global state_generate, index_ex
+    state_generate = True
+    hypothesis = example_data[index_ex]['hypothesis']
+    experiment_plan = example_data[index_ex]['experiment_plan']
+    return hypothesis, experiment_plan, hypothesis, experiment_plan
+
+##########  Phase 2 & 3  ##############
+def start_experiment_agent(hypothesis, plan):
+    if (not state_extract or not state_generate or not state_example):
+        return "", "", ""
+    global state_agent, step_index, state_complete
+    state_agent = True
+    step_index = 0
+    state_complete = False
+    # predefined_message = f"Implement the following hypothesis and experiment plan:\n\nHypothesis:\n{hypothesis}\n\nExperiment Plan:\n{plan}"
+    return example_data[index_ex]['code_init'], predefined_action_log, "", ""
+
+def submit_feedback(user_feedback, history, previous_response):
+    if (not state_extract or not  state_generate or not state_agent or not state_example):
+        return "", "", ""
+    global step_index, state_complete
+    step_index += 1
+    msg = history
+    if step_index < len(process_steps):
+        msg += previous_response + "\nUser feedback:" + user_feedback + "\n\n"
+        response_info = process_steps[step_index]
+        response = info_to_message(response_info)  # Convert dictionary to formatted string
+        response += "Please provide feedback based on the history, response entries, and observation, and questions: "
+        step_index += 1
+        msg += response
+    else:
+        state_complete = True
+        response = "Agent Finished."
+        
+    return msg, response, example_data[index_ex]['code_init'] if state_complete else example_data[index_ex]['code_final'], ""
+
+def load_phase_2_inputs(hypothesis, plan):
+    return hypothesis, plan, "# Code implementation will be displayed here after Start ExperimentAgent."
+
+
+
+predefined_action_log = """
+[Reasoning]: To understand the initial structure and functionality of train.py for effective improvements.
+[Action]: Inspect Script (train.py)
+Input: {"script_name": "train.py", "start_line_number": "1", "end_line_number": "74"}
+Objective: Understand the training script, including data processing, [...]
+[Observation]: The train.py script imports [...]. Sets random seeds [...]. Defines [...] Placeholder functions [...] exist without implementation. [...]
+[Feedback]: The script structure is clear, but key functions (train_model, predict) need proper implementation for proposed model training and prediction.\n
+"""
+
+
+predefined_observation = """
+Epoch [1/10],
+Train MSE: 0.543,
+Test MSE: 0.688
+Epoch [2/10],
+Train MSE: 0.242,
+Test MSE: 0.493\n
+"""
+
+# Initialize the global step_index and history
+process_steps = [
+    {
+        "Action": "Inspect Script Lines (train.py)",
+        "Observation": (
+            "The train.py script imports necessary libraries (e.g., pandas, sklearn, torch). "
+            "Sets random seeds for reproducibility. Defines compute_metrics_for_regression function "
+            "to calculate RMSE for different dimensions. Placeholder functions train_model and "
+            "predict exist without implementations."
+        ),
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The script executed successfully. Generated embeddings using the BERT model. Completed "
+            "the training process without errors. Metrics calculation placeholders indicated areas needing implementation."
+        ),
+    },
+    {
+        "Action": "Edit Script (train.py)",
+        "Observation": (
+            "Edited train.py to separate data loading, model definition, training loop, and evaluation into distinct functions. "
+            "The edited train.py now has clearly defined functions"
+            "for data loading (load_data), model definition (build_model), "
+            "training (train_model), and evaluation (evaluate_model). Similarly, eval.py is reorganized to load the model and perform predictions efficiently."
+        ),
+    },
+    {
+        "Action": "Retrieve Model",
+        "Observation": "CNN and BiLSTM retrieved.",
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The model trained over the specified number of epochs. Training and validation loss values are recorded for each epoch, "
+            "the decrease in loss indicates improved model performance."
+        )
+    },
+    {
+        "Action": "Evaluation",
+        "Observation": predefined_observation,
+    }
+]
+def info_to_message(info):
+    msg = ""
+    for k, v in info.items():
+        if isinstance(v, dict):
+            tempv = v
+            v = ""
+            for k2, v2 in tempv.items():
+                v += f"{k2}:\n  {v2}\n"
+        v = User.indent_text(v, 2)
+        msg += '-' * 64
+        msg += '\n'
+        msg += f"{k}:\n{v}\n"
+    return msg
+
+
+def handle_example_click(example_index):
+    global index_ex
+    index_ex = example_index
+    return load_example(index_ex)  # Simply return the text to display it in the textbox
+
+# Gradio Interface
+with gr.Blocks(css=".gr-examples-label {display: none;}", theme=gr.themes.Default()) as app:
+    gr.Markdown("# MLR- Copilot: Machine Learning Research based on LLM Agents")
+    gr.Markdown("###   ")
+    gr.Markdown("## <span style='color:Orange;'> This UI is for predefined example demo only.</span>")
+    gr.Markdown("## <span style='color:Orange;'> To reproduce the results please use [Github](https://github.com/du-nlp-lab/MLR-Copilot/)</span>")
+
+
+
+    gr.Markdown("MLR-Copilot is a framework where LLMs mimic researchers’ thought processes, designed to enhance the productivity of machine learning research by automating the generation and implementation of research ideas. It begins with a research paper, autonomously generating and validating these ideas, while incorporating human feedback to help reach executable research outcomes.")
+
+
+    gr.Markdown("##    ⬇️  Click an example at bottom to start！")
+
+
+    # Use state variables to store generated hypothesis and experiment plan
+    hypothesis_state = gr.State("")
+    experiment_plan_state = gr.State("")
+
+    ########## Phase 1:  Research Idea Generation Tab ##############
+    with gr.Tab("💡Stage 1: Research Idea Generation"):
+        gr.Markdown("### Extract Research Elements and Generate Research Ideas")
+        with gr.Row():
+            with gr.Column():
+                paper_text_input = gr.Textbox(value="", lines=10, label="📑 Research Paper Text", interactive=False)
+
+                # extract_button = gr.Button("🔍 Extract Research Elements")
+                with gr.Row():
+                    tasks_output = gr.Textbox(placeholder="Research task definition", label="Research Tasks", lines=2, interactive=False)
+                    gaps_output = gr.Textbox(placeholder="Research gaps of current works", label="Research Gaps", lines=2, interactive=False)
+                    keywords_output = gr.Textbox(placeholder="Paper keywords", label="Keywords", lines=2, interactive=False)
+                    recent_works_output = gr.Textbox(placeholder="Recent works extracted from Semantic Scholar", label="Recent Works", lines=2, interactive=False)
+            with gr.Column():
+                with gr.Row():  # Move the button to the top 
+                    generate_button = gr.Button("✍️ Generate Research Hypothesis & Experiment Plan")
+                with gr.Group():
+                    gr.Markdown("### 🌟 Research Idea")
+                    with gr.Row():
+                        hypothesis_output = gr.Textbox(label="Generated Hypothesis", lines=20, interactive=False)
+                        experiment_plan_output = gr.Textbox(label="Generated Experiment Plan", lines=20, interactive=False)
+        
+    
+        # Step 1: Extract Research Elements
+        # extract_button.click(
+        #     fn=extract_research_elements, 
+        #     inputs=paper_text_input, 
+        #     outputs=[tasks_output, gaps_output, keywords_output, recent_works_output]
+        # )
+
+        generate_button.click(
+            fn=generate_and_store, 
+            inputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output], 
+            outputs=[hypothesis_output, experiment_plan_output, hypothesis_state, experiment_plan_state]
+        )
+
+        gr.Examples(
+            examples=example_text,
+            inputs=[paper_text_input],
+            outputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output, hypothesis_output, experiment_plan_output],
+            fn=load_example_and_set_index,
+            run_on_click = True,
+            # label="⬇️ Click an example to load"
+        )
+        
+
+
+
+    ########## Phase 2 & 3:  Experiment implementation and execution ##############
+    with gr.Tab("🧪 Stage 2 & Stage 3: Experiment implementation and execution"):
+        gr.Markdown("### Interact with the ExperimentAgent")
+
+        with gr.Row():
+            with gr.Column():
+                with gr.Group():
+                    gr.Markdown("### 🌟 Generated Research Idea")
+                    with gr.Row():
+                        idea_input = gr.Textbox(label="Generated Research Hypothesis", lines=30, interactive=False)
+                        plan_input = gr.Textbox(label="Generated Experiment Plan", lines=30, interactive=False)
+        
+            with gr.Column():
+                start_exp_agnet = gr.Button("⚙️ Start / Reset ExperimentAgent", elem_classes=["agent-btn"])
+                with gr.Group():
+                    gr.Markdown("### Implementation + Execution Log")    
+                    log = gr.Textbox(label="📖 Execution Log", lines=20, interactive=False)
+                    code_display = gr.Code(label="🧑‍💻 Implementation", language="python", interactive=False)
+
+            with gr.Column():
+                response = gr.Textbox(label="🤖 ExperimentAgent Response", lines=30, interactive=False)
+                feedback = gr.Textbox(placeholder="N/A", label="🧑‍🔬 User Feedback", lines=3, interactive=True)
+                submit_button = gr.Button("Submit", elem_classes=["Submit-btn"])
+        
+        hypothesis_state.change(
+            fn=load_phase_2_inputs, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[idea_input, plan_input, code_display]
+        )
+        
+        # Start research agent
+        start_exp_agnet.click(
+            fn=start_experiment_agent, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[code_display, log, response, feedback]
+        )
+        
+        submit_button.click(
+            fn=submit_feedback,
+            inputs=[feedback, log, response],
+            outputs=[log, response, code_display, feedback]
+        )
+
+# Test
+if __name__ == "__main__":
+    step_index = 0
+    app.launch(share=True)

.history/app_20250404174438.py

@@ -0,0 +1,341 @@
+import gradio as gr
+from pathlib import Path
+from reactagent.environment import Environment
+from reactagent.agents.agent_research import ResearchAgent
+from reactagent.runner import create_parser
+from reactagent import llm
+from reactagent.users.user import User
+import os
+import json
+
+
+# Global variables to store session state
+env = None
+agent = None
+state_example = False
+state_extract = False
+state_generate = False
+state_agent = False
+state_complete = False
+index_ex = "1"
+
+example_text = [
+    "Research Paper 1: Dataset and Baseline for Automatic Student Feedback Analysis",
+    "Research Paper 2: An Empirical Study on the Impact of Code Review on Software Quality"
+]
+
+
+# Load example JSON file
+def load_example_data():
+    with open("example/example_data.json", "r") as json_file:
+        example_data = json.load(json_file)
+    
+    for idx in example_data.keys():
+        try:
+            file = example_data[idx]["code_init"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_init"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")
+        try:
+            file = example_data[idx]["code_final"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_final"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")     
+    return example_data
+
+example_data = load_example_data()
+
+# Function to handle the selection of an example and populate the respective fields
+def load_example(example_id):
+    global index_ex 
+    index_ex = str(example_id)
+    example = example_data[index_ex]
+    paper_text = 'Title:\t' +  example['title'] + '\n\nAbstract:\t' + example['abstract']
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return [paper_text, tasks, gaps, keywords, recent_works]
+
+# example_text = [load_example(1), load_example(2)]
+# # WRONG: example_text = [load_example(1), load_example(2)]
+
+# ✅ FIXED:
+example_text = []
+for i in ["1", "2"]:
+    paper_text = 'Title:\t' + example_data[i]['title'] + '\n\nAbstract:\t' + example_data[i]['abstract']
+    example_text.append(paper_text)
+
+# Function to handle example clicks
+def load_example_and_set_index(paper_text_input):
+    global index_ex, state_example
+    state_example = True
+    index_ex = str(example_text.index(paper_text_input) + 1)
+    example =  load_example(index_ex)
+
+    return example
+
+
+
+########## Phase 1 ##############
+
+def extract_research_elements(paper_text):
+    global state_extract, index_ex, state_example
+    if not state_example or paper_text == "":
+        return "", "", "", ""
+    state_extract = True
+    if not paper_text.strip().startswith("Title:\t" + example_data[index_ex]["title"]):
+        print("Mismatch detected.")
+        print(paper_text)
+        return "", "", "", ""
+    example = example_data[index_ex]
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return tasks, gaps, keywords, recent_works
+
+
+# Step 2: Generate Research Hypothesis and Experiment Plan
+def generate_and_store(paper_text, tasks, gaps, keywords, recent_works):
+    if (not state_extract or not state_example or paper_text == ""):
+        return "", "", "", ""
+    global state_generate, index_ex
+    state_generate = True
+    hypothesis = example_data[index_ex]['hypothesis']
+    experiment_plan = example_data[index_ex]['experiment_plan']
+    return hypothesis, experiment_plan, hypothesis, experiment_plan
+
+##########  Phase 2 & 3  ##############
+def start_experiment_agent(hypothesis, plan):
+    if (not state_extract or not state_generate or not state_example):
+        return "", "", ""
+    global state_agent, step_index, state_complete
+    state_agent = True
+    step_index = 0
+    state_complete = False
+    # predefined_message = f"Implement the following hypothesis and experiment plan:\n\nHypothesis:\n{hypothesis}\n\nExperiment Plan:\n{plan}"
+    return example_data[index_ex]['code_init'], predefined_action_log, "", ""
+
+def submit_feedback(user_feedback, history, previous_response):
+    if (not state_extract or not  state_generate or not state_agent or not state_example):
+        return "", "", ""
+    global step_index, state_complete
+    step_index += 1
+    msg = history
+    if step_index < len(process_steps):
+        msg += previous_response + "\nUser feedback:" + user_feedback + "\n\n"
+        response_info = process_steps[step_index]
+        response = info_to_message(response_info)  # Convert dictionary to formatted string
+        response += "Please provide feedback based on the history, response entries, and observation, and questions: "
+        step_index += 1
+        msg += response
+    else:
+        state_complete = True
+        response = "Agent Finished."
+        
+    return msg, response, example_data[index_ex]['code_init'] if state_complete else example_data[index_ex]['code_final'], ""
+
+def load_phase_2_inputs(hypothesis, plan):
+    return hypothesis, plan, "# Code implementation will be displayed here after Start ExperimentAgent."
+
+
+
+predefined_action_log = """
+[Reasoning]: To understand the initial structure and functionality of train.py for effective improvements.
+[Action]: Inspect Script (train.py)
+Input: {"script_name": "train.py", "start_line_number": "1", "end_line_number": "74"}
+Objective: Understand the training script, including data processing, [...]
+[Observation]: The train.py script imports [...]. Sets random seeds [...]. Defines [...] Placeholder functions [...] exist without implementation. [...]
+[Feedback]: The script structure is clear, but key functions (train_model, predict) need proper implementation for proposed model training and prediction.\n
+"""
+
+
+predefined_observation = """
+Epoch [1/10],
+Train MSE: 0.543,
+Test MSE: 0.688
+Epoch [2/10],
+Train MSE: 0.242,
+Test MSE: 0.493\n
+"""
+
+# Initialize the global step_index and history
+process_steps = [
+    {
+        "Action": "Inspect Script Lines (train.py)",
+        "Observation": (
+            "The train.py script imports necessary libraries (e.g., pandas, sklearn, torch). "
+            "Sets random seeds for reproducibility. Defines compute_metrics_for_regression function "
+            "to calculate RMSE for different dimensions. Placeholder functions train_model and "
+            "predict exist without implementations."
+        ),
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The script executed successfully. Generated embeddings using the BERT model. Completed "
+            "the training process without errors. Metrics calculation placeholders indicated areas needing implementation."
+        ),
+    },
+    {
+        "Action": "Edit Script (train.py)",
+        "Observation": (
+            "Edited train.py to separate data loading, model definition, training loop, and evaluation into distinct functions. "
+            "The edited train.py now has clearly defined functions"
+            "for data loading (load_data), model definition (build_model), "
+            "training (train_model), and evaluation (evaluate_model). Similarly, eval.py is reorganized to load the model and perform predictions efficiently."
+        ),
+    },
+    {
+        "Action": "Retrieve Model",
+        "Observation": "CNN and BiLSTM retrieved.",
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The model trained over the specified number of epochs. Training and validation loss values are recorded for each epoch, "
+            "the decrease in loss indicates improved model performance."
+        )
+    },
+    {
+        "Action": "Evaluation",
+        "Observation": predefined_observation,
+    }
+]
+def info_to_message(info):
+    msg = ""
+    for k, v in info.items():
+        if isinstance(v, dict):
+            tempv = v
+            v = ""
+            for k2, v2 in tempv.items():
+                v += f"{k2}:\n  {v2}\n"
+        v = User.indent_text(v, 2)
+        msg += '-' * 64
+        msg += '\n'
+        msg += f"{k}:\n{v}\n"
+    return msg
+
+
+def handle_example_click(example_index):
+    global index_ex
+    index_ex = example_index
+    return load_example(index_ex)  # Simply return the text to display it in the textbox
+
+# Gradio Interface
+with gr.Blocks(css=".gr-examples-label {display: none;}", theme=gr.themes.Default()) as app:
+    gr.Markdown("# MLR- Copilot: Machine Learning Research based on LLM Agents")
+    gr.Markdown("###   ")
+    gr.Markdown("## <span style='color:Orange;'> This UI is for predefined example demo only.</span>")
+    gr.Markdown("## <span style='color:Orange;'> To reproduce the results please use [Github](https://github.com/du-nlp-lab/MLR-Copilot/)</span>")
+
+
+
+    gr.Markdown("MLR-Copilot is a framework where LLMs mimic researchers’ thought processes, designed to enhance the productivity of machine learning research by automating the generation and implementation of research ideas. It begins with a research paper, autonomously generating and validating these ideas, while incorporating human feedback to help reach executable research outcomes.")
+
+
+    gr.Markdown("##    ⬇️  Click an example at bottom to start！")
+
+
+    # Use state variables to store generated hypothesis and experiment plan
+    hypothesis_state = gr.State("")
+    experiment_plan_state = gr.State("")
+
+    ########## Phase 1:  Research Idea Generation Tab ##############
+    with gr.Tab("💡Stage 1: Research Idea Generation"):
+        gr.Markdown("### Extract Research Elements and Generate Research Ideas")
+        with gr.Row():
+            with gr.Column():
+                paper_text_input = gr.Textbox(value="", lines=10, label="📑 Research Paper Text", interactive=False)
+
+                # extract_button = gr.Button("🔍 Extract Research Elements")
+                with gr.Row():
+                    tasks_output = gr.Textbox(placeholder="Research task definition", label="Research Tasks", lines=2, interactive=False)
+                    gaps_output = gr.Textbox(placeholder="Research gaps of current works", label="Research Gaps", lines=2, interactive=False)
+                    keywords_output = gr.Textbox(placeholder="Paper keywords", label="Keywords", lines=2, interactive=False)
+                    recent_works_output = gr.Textbox(placeholder="Recent works extracted from Semantic Scholar", label="Recent Works", lines=2, interactive=False)
+            with gr.Column():
+                with gr.Row():  # Move the button to the top 
+                    generate_button = gr.Button("✍️ Generate Research Hypothesis & Experiment Plan")
+                with gr.Group():
+                    gr.Markdown("### 🌟 Research Idea")
+                    with gr.Row():
+                        hypothesis_output = gr.Textbox(label="Generated Hypothesis", lines=20, interactive=False)
+                        experiment_plan_output = gr.Textbox(label="Generated Experiment Plan", lines=20, interactive=False)
+        
+    
+        # Step 1: Extract Research Elements
+        # extract_button.click(
+        #     fn=extract_research_elements, 
+        #     inputs=paper_text_input, 
+        #     outputs=[tasks_output, gaps_output, keywords_output, recent_works_output]
+        # )
+
+        generate_button.click(
+            fn=generate_and_store, 
+            inputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output], 
+            outputs=[hypothesis_output, experiment_plan_output, hypothesis_state, experiment_plan_state]
+        )
+
+        gr.Examples(
+            examples=example_text,
+            inputs=[paper_text_input],
+            outputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output, hypothesis_output, experiment_plan_output],
+            fn=load_example_and_set_index,
+            run_on_click = True,
+            # label="⬇️ Click an example to load"
+        )
+        
+
+
+
+    ########## Phase 2 & 3:  Experiment implementation and execution ##############
+    with gr.Tab("🧪 Stage 2 & Stage 3: Experiment implementation and execution"):
+        gr.Markdown("### Interact with the ExperimentAgent")
+
+        with gr.Row():
+            with gr.Column():
+                with gr.Group():
+                    gr.Markdown("### 🌟 Generated Research Idea")
+                    with gr.Row():
+                        idea_input = gr.Textbox(label="Generated Research Hypothesis", lines=30, interactive=False)
+                        plan_input = gr.Textbox(label="Generated Experiment Plan", lines=30, interactive=False)
+        
+            with gr.Column():
+                start_exp_agnet = gr.Button("⚙️ Start / Reset ExperimentAgent", elem_classes=["agent-btn"])
+                with gr.Group():
+                    gr.Markdown("### Implementation + Execution Log")    
+                    log = gr.Textbox(label="📖 Execution Log", lines=20, interactive=False)
+                    code_display = gr.Code(label="🧑‍💻 Implementation", language="python", interactive=False)
+
+            with gr.Column():
+                response = gr.Textbox(label="🤖 ExperimentAgent Response", lines=30, interactive=False)
+                feedback = gr.Textbox(placeholder="N/A", label="🧑‍🔬 User Feedback", lines=3, interactive=True)
+                submit_button = gr.Button("Submit", elem_classes=["Submit-btn"])
+        
+        hypothesis_state.change(
+            fn=load_phase_2_inputs, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[idea_input, plan_input, code_display]
+        )
+        
+        # Start research agent
+        start_exp_agnet.click(
+            fn=start_experiment_agent, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[code_display, log, response, feedback]
+        )
+        
+        submit_button.click(
+            fn=submit_feedback,
+            inputs=[feedback, log, response],
+            outputs=[log, response, code_display, feedback]
+        )
+
+# Test
+if __name__ == "__main__":
+    step_index = 0
+    app.launch(share=True)

.history/app_20250404174439.py

@@ -0,0 +1,341 @@
+import gradio as gr
+from pathlib import Path
+from reactagent.environment import Environment
+from reactagent.agents.agent_research import ResearchAgent
+from reactagent.runner import create_parser
+from reactagent import llm
+from reactagent.users.user import User
+import os
+import json
+
+
+# Global variables to store session state
+env = None
+agent = None
+state_example = False
+state_extract = False
+state_generate = False
+state_agent = False
+state_complete = False
+index_ex = "1"
+
+example_text = [
+    "Research Paper 1: Dataset and Baseline for Automatic Student Feedback Analysis",
+    "Research Paper 2: An Empirical Study on the Impact of Code Review on Software Quality"
+]
+
+
+# Load example JSON file
+def load_example_data():
+    with open("example/example_data.json", "r") as json_file:
+        example_data = json.load(json_file)
+    
+    for idx in example_data.keys():
+        try:
+            file = example_data[idx]["code_init"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_init"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")
+        try:
+            file = example_data[idx]["code_final"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_final"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")     
+    return example_data
+
+example_data = load_example_data()
+
+# Function to handle the selection of an example and populate the respective fields
+def load_example(example_id):
+    global index_ex 
+    index_ex = str(example_id)
+    example = example_data[index_ex]
+    paper_text = 'Title:\t' +  example['title'] + '\n\nAbstract:\t' + example['abstract']
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return [paper_text, tasks, gaps, keywords, recent_works]
+
+# example_text = [load_example(1), load_example(2)]
+# # WRONG: example_text = [load_example(1), load_example(2)]
+
+# ✅ FIXED:
+example_text = []
+for i in ["1", "2"]:
+    paper_text = 'Title:\t' + example_data[i]['title'] + '\n\nAbstract:\t' + example_data[i]['abstract']
+    example_text.append(paper_text)
+
+# Function to handle example clicks
+def load_example_and_set_index(paper_text_input):
+    global index_ex, state_example
+    state_example = True
+    index_ex = str(example_text.index(paper_text_input) + 1)
+    example =  load_example(index_ex)
+
+    return example
+
+
+
+########## Phase 1 ##############
+
+def extract_research_elements(paper_text):
+    global state_extract, index_ex, state_example
+    if not state_example or paper_text == "":
+        return "", "", "", ""
+    state_extract = True
+    if not paper_text.strip().startswith("Title:\t" + example_data[index_ex]["title"]):
+        print("Mismatch detected.")
+        print(paper_text)
+        return "", "", "", ""
+    example = example_data[index_ex]
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return tasks, gaps, keywords, recent_works
+
+
+# Step 2: Generate Research Hypothesis and Experiment Plan
+def generate_and_store(paper_text, tasks, gaps, keywords, recent_works):
+    if (not state_extract or not state_example or paper_text == ""):
+        return "", "", "", ""
+    global state_generate, index_ex
+    state_generate = True
+    hypothesis = example_data[index_ex]['hypothesis']
+    experiment_plan = example_data[index_ex]['experiment_plan']
+    return hypothesis, experiment_plan, hypothesis, experiment_plan
+
+##########  Phase 2 & 3  ##############
+def start_experiment_agent(hypothesis, plan):
+    if (not state_extract or not state_generate or not state_example):
+        return "", "", ""
+    global state_agent, step_index, state_complete
+    state_agent = True
+    step_index = 0
+    state_complete = False
+    # predefined_message = f"Implement the following hypothesis and experiment plan:\n\nHypothesis:\n{hypothesis}\n\nExperiment Plan:\n{plan}"
+    return example_data[index_ex]['code_init'], predefined_action_log, "", ""
+
+def submit_feedback(user_feedback, history, previous_response):
+    if (not state_extract or not  state_generate or not state_agent or not state_example):
+        return "", "", ""
+    global step_index, state_complete
+    step_index += 1
+    msg = history
+    if step_index < len(process_steps):
+        msg += previous_response + "\nUser feedback:" + user_feedback + "\n\n"
+        response_info = process_steps[step_index]
+        response = info_to_message(response_info)  # Convert dictionary to formatted string
+        response += "Please provide feedback based on the history, response entries, and observation, and questions: "
+        step_index += 1
+        msg += response
+    else:
+        state_complete = True
+        response = "Agent Finished."
+        
+    return msg, response, example_data[index_ex]['code_init'] if state_complete else example_data[index_ex]['code_final'], ""
+
+def load_phase_2_inputs(hypothesis, plan):
+    return hypothesis, plan, "# Code implementation will be displayed here after Start ExperimentAgent."
+
+
+
+predefined_action_log = """
+[Reasoning]: To understand the initial structure and functionality of train.py for effective improvements.
+[Action]: Inspect Script (train.py)
+Input: {"script_name": "train.py", "start_line_number": "1", "end_line_number": "74"}
+Objective: Understand the training script, including data processing, [...]
+[Observation]: The train.py script imports [...]. Sets random seeds [...]. Defines [...] Placeholder functions [...] exist without implementation. [...]
+[Feedback]: The script structure is clear, but key functions (train_model, predict) need proper implementation for proposed model training and prediction.\n
+"""
+
+
+predefined_observation = """
+Epoch [1/10],
+Train MSE: 0.543,
+Test MSE: 0.688
+Epoch [2/10],
+Train MSE: 0.242,
+Test MSE: 0.493\n
+"""
+
+# Initialize the global step_index and history
+process_steps = [
+    {
+        "Action": "Inspect Script Lines (train.py)",
+        "Observation": (
+            "The train.py script imports necessary libraries (e.g., pandas, sklearn, torch). "
+            "Sets random seeds for reproducibility. Defines compute_metrics_for_regression function "
+            "to calculate RMSE for different dimensions. Placeholder functions train_model and "
+            "predict exist without implementations."
+        ),
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The script executed successfully. Generated embeddings using the BERT model. Completed "
+            "the training process without errors. Metrics calculation placeholders indicated areas needing implementation."
+        ),
+    },
+    {
+        "Action": "Edit Script (train.py)",
+        "Observation": (
+            "Edited train.py to separate data loading, model definition, training loop, and evaluation into distinct functions. "
+            "The edited train.py now has clearly defined functions"
+            "for data loading (load_data), model definition (build_model), "
+            "training (train_model), and evaluation (evaluate_model). Similarly, eval.py is reorganized to load the model and perform predictions efficiently."
+        ),
+    },
+    {
+        "Action": "Retrieve Model",
+        "Observation": "CNN and BiLSTM retrieved.",
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The model trained over the specified number of epochs. Training and validation loss values are recorded for each epoch, "
+            "the decrease in loss indicates improved model performance."
+        )
+    },
+    {
+        "Action": "Evaluation",
+        "Observation": predefined_observation,
+    }
+]
+def info_to_message(info):
+    msg = ""
+    for k, v in info.items():
+        if isinstance(v, dict):
+            tempv = v
+            v = ""
+            for k2, v2 in tempv.items():
+                v += f"{k2}:\n  {v2}\n"
+        v = User.indent_text(v, 2)
+        msg += '-' * 64
+        msg += '\n'
+        msg += f"{k}:\n{v}\n"
+    return msg
+
+
+def handle_example_click(example_index):
+    global index_ex
+    index_ex = example_index
+    return load_example(index_ex)  # Simply return the text to display it in the textbox
+
+# Gradio Interface
+with gr.Blocks(css=".gr-examples-label {display: none;}", theme=gr.themes.Default()) as app:
+    gr.Markdown("# MLR- Copilot: Machine Learning Research based on LLM Agents")
+    gr.Markdown("###   ")
+    gr.Markdown("## <span style='color:Orange;'> This UI is for predefined example demo only.</span>")
+    gr.Markdown("## <span style='color:Orange;'> To reproduce the results please use [Github](https://github.com/du-nlp-lab/MLR-Copilot/)</span>")
+
+
+
+    gr.Markdown("MLR-Copilot is a framework where LLMs mimic researchers’ thought processes, designed to enhance the productivity of machine learning research by automating the generation and implementation of research ideas. It begins with a research paper, autonomously generating and validating these ideas, while incorporating human feedback to help reach executable research outcomes.")
+
+
+    gr.Markdown("##    ⬇️  Click an example at bottom to start！")
+
+
+    # Use state variables to store generated hypothesis and experiment plan
+    hypothesis_state = gr.State("")
+    experiment_plan_state = gr.State("")
+
+    ########## Phase 1:  Research Idea Generation Tab ##############
+    with gr.Tab("💡Stage 1: Research Idea Generation"):
+        gr.Markdown("### Extract Research Elements and Generate Research Ideas")
+        with gr.Row():
+            with gr.Column():
+                paper_text_input = gr.Textbox(value="", lines=10, label="📑 Research Paper Text", interactive=False)
+
+                # extract_button = gr.Button("🔍 Extract Research Elements")
+                with gr.Row():
+                    tasks_output = gr.Textbox(placeholder="Research task definition", label="Research Tasks", lines=2, interactive=False)
+                    gaps_output = gr.Textbox(placeholder="Research gaps of current works", label="Research Gaps", lines=2, interactive=False)
+                    keywords_output = gr.Textbox(placeholder="Paper keywords", label="Keywords", lines=2, interactive=False)
+                    recent_works_output = gr.Textbox(placeholder="Recent works extracted from Semantic Scholar", label="Recent Works", lines=2, interactive=False)
+            with gr.Column():
+                with gr.Row():  # Move the button to the top 
+                    generate_button = gr.Button("✍️ Generate Research Hypothesis & Experiment Plan")
+                with gr.Group():
+                    gr.Markdown("### 🌟 Research Idea")
+                    with gr.Row():
+                        hypothesis_output = gr.Textbox(label="Generated Hypothesis", lines=20, interactive=False)
+                        experiment_plan_output = gr.Textbox(label="Generated Experiment Plan", lines=20, interactive=False)
+        
+    
+        # Step 1: Extract Research Elements
+        # extract_button.click(
+        #     fn=extract_research_elements, 
+        #     inputs=paper_text_input, 
+        #     outputs=[tasks_output, gaps_output, keywords_output, recent_works_output]
+        # )
+
+        generate_button.click(
+            fn=generate_and_store, 
+            inputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output], 
+            outputs=[hypothesis_output, experiment_plan_output, hypothesis_state, experiment_plan_state]
+        )
+
+        gr.Examples(
+            examples=example_text,
+            inputs=[paper_text_input],
+            outputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output, hypothesis_output, experiment_plan_output],
+            fn=load_example_and_set_index,
+            run_on_click = True,
+            # label="⬇️ Click an example to load"
+        )
+        
+
+
+
+    ########## Phase 2 & 3:  Experiment implementation and execution ##############
+    with gr.Tab("🧪 Stage 2 & Stage 3: Experiment implementation and execution"):
+        gr.Markdown("### Interact with the ExperimentAgent")
+
+        with gr.Row():
+            with gr.Column():
+                with gr.Group():
+                    gr.Markdown("### 🌟 Generated Research Idea")
+                    with gr.Row():
+                        idea_input = gr.Textbox(label="Generated Research Hypothesis", lines=30, interactive=False)
+                        plan_input = gr.Textbox(label="Generated Experiment Plan", lines=30, interactive=False)
+        
+            with gr.Column():
+                start_exp_agnet = gr.Button("⚙️ Start / Reset ExperimentAgent", elem_classes=["agent-btn"])
+                with gr.Group():
+                    gr.Markdown("### Implementation + Execution Log")    
+                    log = gr.Textbox(label="📖 Execution Log", lines=20, interactive=False)
+                    code_display = gr.Code(label="🧑‍💻 Implementation", language="python", interactive=False)
+
+            with gr.Column():
+                response = gr.Textbox(label="🤖 ExperimentAgent Response", lines=30, interactive=False)
+                feedback = gr.Textbox(placeholder="N/A", label="🧑‍🔬 User Feedback", lines=3, interactive=True)
+                submit_button = gr.Button("Submit", elem_classes=["Submit-btn"])
+        
+        hypothesis_state.change(
+            fn=load_phase_2_inputs, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[idea_input, plan_input, code_display]
+        )
+        
+        # Start research agent
+        start_exp_agnet.click(
+            fn=start_experiment_agent, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[code_display, log, response, feedback]
+        )
+        
+        submit_button.click(
+            fn=submit_feedback,
+            inputs=[feedback, log, response],
+            outputs=[log, response, code_display, feedback]
+        )
+
+# Test
+if __name__ == "__main__":
+    step_index = 0
+    app.launch(share=True)

.history/app_20250404174544.py

@@ -0,0 +1,335 @@
+import gradio as gr
+from pathlib import Path
+from reactagent.environment import Environment
+from reactagent.agents.agent_research import ResearchAgent
+from reactagent.runner import create_parser
+from reactagent import llm
+from reactagent.users.user import User
+import os
+import json
+
+
+# Global variables to store session state
+env = None
+agent = None
+state_example = False
+state_extract = False
+state_generate = False
+state_agent = False
+state_complete = False
+index_ex = "1"
+
+example_text = [
+    "Research Paper 1: Dataset and Baseline for Automatic Student Feedback Analysis",
+    "Research Paper 2: An Empirical Study on the Impact of Code Review on Software Quality"
+]
+
+
+# Load example JSON file
+def load_example_data():
+    with open("example/example_data.json", "r") as json_file:
+        example_data = json.load(json_file)
+    
+    for idx in example_data.keys():
+        try:
+            file = example_data[idx]["code_init"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_init"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")
+        try:
+            file = example_data[idx]["code_final"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_final"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")     
+    return example_data
+
+example_data = load_example_data()
+
+# Function to handle the selection of an example and populate the respective fields
+def load_example(example_id):
+    global index_ex 
+    index_ex = str(example_id)
+    example = example_data[index_ex]
+    paper_text = 'Title:\t' +  example['title'] + '\n\nAbstract:\t' + example['abstract']
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return [paper_text, tasks, gaps, keywords, recent_works]
+
+example_text = [load_example(1)[0], load_example(2)[0]]
+
+
+# Function to handle example clicks
+def load_example_and_set_index(paper_text_input):
+    global index_ex, state_example
+    state_example = True
+    index_ex = str(example_text.index(paper_text_input) + 1)
+    example =  load_example(index_ex)
+
+    return example
+
+
+
+########## Phase 1 ##############
+
+def extract_research_elements(paper_text):
+    global state_extract, index_ex, state_example
+    if not state_example or paper_text == "":
+        return "", "", "", ""
+    state_extract = True
+    if not paper_text.strip().startswith("Title:\t" + example_data[index_ex]["title"]):
+        print("Mismatch detected.")
+        print(paper_text)
+        return "", "", "", ""
+    example = example_data[index_ex]
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return tasks, gaps, keywords, recent_works
+
+
+# Step 2: Generate Research Hypothesis and Experiment Plan
+def generate_and_store(paper_text, tasks, gaps, keywords, recent_works):
+    if (not state_extract or not state_example or paper_text == ""):
+        return "", "", "", ""
+    global state_generate, index_ex
+    state_generate = True
+    hypothesis = example_data[index_ex]['hypothesis']
+    experiment_plan = example_data[index_ex]['experiment_plan']
+    return hypothesis, experiment_plan, hypothesis, experiment_plan
+
+##########  Phase 2 & 3  ##############
+def start_experiment_agent(hypothesis, plan):
+    if (not state_extract or not state_generate or not state_example):
+        return "", "", ""
+    global state_agent, step_index, state_complete
+    state_agent = True
+    step_index = 0
+    state_complete = False
+    # predefined_message = f"Implement the following hypothesis and experiment plan:\n\nHypothesis:\n{hypothesis}\n\nExperiment Plan:\n{plan}"
+    return example_data[index_ex]['code_init'], predefined_action_log, "", ""
+
+def submit_feedback(user_feedback, history, previous_response):
+    if (not state_extract or not  state_generate or not state_agent or not state_example):
+        return "", "", ""
+    global step_index, state_complete
+    step_index += 1
+    msg = history
+    if step_index < len(process_steps):
+        msg += previous_response + "\nUser feedback:" + user_feedback + "\n\n"
+        response_info = process_steps[step_index]
+        response = info_to_message(response_info)  # Convert dictionary to formatted string
+        response += "Please provide feedback based on the history, response entries, and observation, and questions: "
+        step_index += 1
+        msg += response
+    else:
+        state_complete = True
+        response = "Agent Finished."
+        
+    return msg, response, example_data[index_ex]['code_init'] if state_complete else example_data[index_ex]['code_final'], ""
+
+def load_phase_2_inputs(hypothesis, plan):
+    return hypothesis, plan, "# Code implementation will be displayed here after Start ExperimentAgent."
+
+
+
+predefined_action_log = """
+[Reasoning]: To understand the initial structure and functionality of train.py for effective improvements.
+[Action]: Inspect Script (train.py)
+Input: {"script_name": "train.py", "start_line_number": "1", "end_line_number": "74"}
+Objective: Understand the training script, including data processing, [...]
+[Observation]: The train.py script imports [...]. Sets random seeds [...]. Defines [...] Placeholder functions [...] exist without implementation. [...]
+[Feedback]: The script structure is clear, but key functions (train_model, predict) need proper implementation for proposed model training and prediction.\n
+"""
+
+
+predefined_observation = """
+Epoch [1/10],
+Train MSE: 0.543,
+Test MSE: 0.688
+Epoch [2/10],
+Train MSE: 0.242,
+Test MSE: 0.493\n
+"""
+
+# Initialize the global step_index and history
+process_steps = [
+    {
+        "Action": "Inspect Script Lines (train.py)",
+        "Observation": (
+            "The train.py script imports necessary libraries (e.g., pandas, sklearn, torch). "
+            "Sets random seeds for reproducibility. Defines compute_metrics_for_regression function "
+            "to calculate RMSE for different dimensions. Placeholder functions train_model and "
+            "predict exist without implementations."
+        ),
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The script executed successfully. Generated embeddings using the BERT model. Completed "
+            "the training process without errors. Metrics calculation placeholders indicated areas needing implementation."
+        ),
+    },
+    {
+        "Action": "Edit Script (train.py)",
+        "Observation": (
+            "Edited train.py to separate data loading, model definition, training loop, and evaluation into distinct functions. "
+            "The edited train.py now has clearly defined functions"
+            "for data loading (load_data), model definition (build_model), "
+            "training (train_model), and evaluation (evaluate_model). Similarly, eval.py is reorganized to load the model and perform predictions efficiently."
+        ),
+    },
+    {
+        "Action": "Retrieve Model",
+        "Observation": "CNN and BiLSTM retrieved.",
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The model trained over the specified number of epochs. Training and validation loss values are recorded for each epoch, "
+            "the decrease in loss indicates improved model performance."
+        )
+    },
+    {
+        "Action": "Evaluation",
+        "Observation": predefined_observation,
+    }
+]
+def info_to_message(info):
+    msg = ""
+    for k, v in info.items():
+        if isinstance(v, dict):
+            tempv = v
+            v = ""
+            for k2, v2 in tempv.items():
+                v += f"{k2}:\n  {v2}\n"
+        v = User.indent_text(v, 2)
+        msg += '-' * 64
+        msg += '\n'
+        msg += f"{k}:\n{v}\n"
+    return msg
+
+
+def handle_example_click(example_index):
+    global index_ex
+    index_ex = example_index
+    return load_example(index_ex)  # Simply return the text to display it in the textbox
+
+# Gradio Interface
+with gr.Blocks(css=".gr-examples-label {display: none;}", theme=gr.themes.Default()) as app:
+    gr.Markdown("# MLR- Copilot: Machine Learning Research based on LLM Agents")
+    gr.Markdown("###   ")
+    gr.Markdown("## <span style='color:Orange;'> This UI is for predefined example demo only.</span>")
+    gr.Markdown("## <span style='color:Orange;'> To reproduce the results please use [Github](https://github.com/du-nlp-lab/MLR-Copilot/)</span>")
+
+
+
+    gr.Markdown("MLR-Copilot is a framework where LLMs mimic researchers’ thought processes, designed to enhance the productivity of machine learning research by automating the generation and implementation of research ideas. It begins with a research paper, autonomously generating and validating these ideas, while incorporating human feedback to help reach executable research outcomes.")
+
+
+    gr.Markdown("##    ⬇️  Click an example at bottom to start！")
+
+
+    # Use state variables to store generated hypothesis and experiment plan
+    hypothesis_state = gr.State("")
+    experiment_plan_state = gr.State("")
+
+    ########## Phase 1:  Research Idea Generation Tab ##############
+    with gr.Tab("💡Stage 1: Research Idea Generation"):
+        gr.Markdown("### Extract Research Elements and Generate Research Ideas")
+        with gr.Row():
+            with gr.Column():
+                paper_text_input = gr.Textbox(value="", lines=10, label="📑 Research Paper Text", interactive=False)
+
+                # extract_button = gr.Button("🔍 Extract Research Elements")
+                with gr.Row():
+                    tasks_output = gr.Textbox(placeholder="Research task definition", label="Research Tasks", lines=2, interactive=False)
+                    gaps_output = gr.Textbox(placeholder="Research gaps of current works", label="Research Gaps", lines=2, interactive=False)
+                    keywords_output = gr.Textbox(placeholder="Paper keywords", label="Keywords", lines=2, interactive=False)
+                    recent_works_output = gr.Textbox(placeholder="Recent works extracted from Semantic Scholar", label="Recent Works", lines=2, interactive=False)
+            with gr.Column():
+                with gr.Row():  # Move the button to the top 
+                    generate_button = gr.Button("✍️ Generate Research Hypothesis & Experiment Plan")
+                with gr.Group():
+                    gr.Markdown("### 🌟 Research Idea")
+                    with gr.Row():
+                        hypothesis_output = gr.Textbox(label="Generated Hypothesis", lines=20, interactive=False)
+                        experiment_plan_output = gr.Textbox(label="Generated Experiment Plan", lines=20, interactive=False)
+        
+    
+        # Step 1: Extract Research Elements
+        # extract_button.click(
+        #     fn=extract_research_elements, 
+        #     inputs=paper_text_input, 
+        #     outputs=[tasks_output, gaps_output, keywords_output, recent_works_output]
+        # )
+
+        generate_button.click(
+            fn=generate_and_store, 
+            inputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output], 
+            outputs=[hypothesis_output, experiment_plan_output, hypothesis_state, experiment_plan_state]
+        )
+
+        gr.Examples(
+            examples=example_text,
+            inputs=[paper_text_input],
+            outputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output, hypothesis_output, experiment_plan_output],
+            fn=load_example_and_set_index,
+            run_on_click = True,
+            # label="⬇️ Click an example to load"
+        )
+        
+
+
+
+    ########## Phase 2 & 3:  Experiment implementation and execution ##############
+    with gr.Tab("🧪 Stage 2 & Stage 3: Experiment implementation and execution"):
+        gr.Markdown("### Interact with the ExperimentAgent")
+
+        with gr.Row():
+            with gr.Column():
+                with gr.Group():
+                    gr.Markdown("### 🌟 Generated Research Idea")
+                    with gr.Row():
+                        idea_input = gr.Textbox(label="Generated Research Hypothesis", lines=30, interactive=False)
+                        plan_input = gr.Textbox(label="Generated Experiment Plan", lines=30, interactive=False)
+        
+            with gr.Column():
+                start_exp_agnet = gr.Button("⚙️ Start / Reset ExperimentAgent", elem_classes=["agent-btn"])
+                with gr.Group():
+                    gr.Markdown("### Implementation + Execution Log")    
+                    log = gr.Textbox(label="📖 Execution Log", lines=20, interactive=False)
+                    code_display = gr.Code(label="🧑‍💻 Implementation", language="python", interactive=False)
+
+            with gr.Column():
+                response = gr.Textbox(label="🤖 ExperimentAgent Response", lines=30, interactive=False)
+                feedback = gr.Textbox(placeholder="N/A", label="🧑‍🔬 User Feedback", lines=3, interactive=True)
+                submit_button = gr.Button("Submit", elem_classes=["Submit-btn"])
+        
+        hypothesis_state.change(
+            fn=load_phase_2_inputs, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[idea_input, plan_input, code_display]
+        )
+        
+        # Start research agent
+        start_exp_agnet.click(
+            fn=start_experiment_agent, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[code_display, log, response, feedback]
+        )
+        
+        submit_button.click(
+            fn=submit_feedback,
+            inputs=[feedback, log, response],
+            outputs=[log, response, code_display, feedback]
+        )
+
+# Test
+if __name__ == "__main__":
+    step_index = 0
+    app.launch(share=True)

.history/app_20250404174545.py

@@ -0,0 +1,335 @@
+import gradio as gr
+from pathlib import Path
+from reactagent.environment import Environment
+from reactagent.agents.agent_research import ResearchAgent
+from reactagent.runner import create_parser
+from reactagent import llm
+from reactagent.users.user import User
+import os
+import json
+
+
+# Global variables to store session state
+env = None
+agent = None
+state_example = False
+state_extract = False
+state_generate = False
+state_agent = False
+state_complete = False
+index_ex = "1"
+
+example_text = [
+    "Research Paper 1: Dataset and Baseline for Automatic Student Feedback Analysis",
+    "Research Paper 2: An Empirical Study on the Impact of Code Review on Software Quality"
+]
+
+
+# Load example JSON file
+def load_example_data():
+    with open("example/example_data.json", "r") as json_file:
+        example_data = json.load(json_file)
+    
+    for idx in example_data.keys():
+        try:
+            file = example_data[idx]["code_init"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_init"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")
+        try:
+            file = example_data[idx]["code_final"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_final"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")     
+    return example_data
+
+example_data = load_example_data()
+
+# Function to handle the selection of an example and populate the respective fields
+def load_example(example_id):
+    global index_ex 
+    index_ex = str(example_id)
+    example = example_data[index_ex]
+    paper_text = 'Title:\t' +  example['title'] + '\n\nAbstract:\t' + example['abstract']
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return [paper_text, tasks, gaps, keywords, recent_works]
+
+example_text = [load_example(1)[0], load_example(2)[0]]
+
+
+# Function to handle example clicks
+def load_example_and_set_index(paper_text_input):
+    global index_ex, state_example
+    state_example = True
+    index_ex = str(example_text.index(paper_text_input) + 1)
+    example =  load_example(index_ex)
+
+    return example
+
+
+
+########## Phase 1 ##############
+
+def extract_research_elements(paper_text):
+    global state_extract, index_ex, state_example
+    if not state_example or paper_text == "":
+        return "", "", "", ""
+    state_extract = True
+    if not paper_text.strip().startswith("Title:\t" + example_data[index_ex]["title"]):
+        print("Mismatch detected.")
+        print(paper_text)
+        return "", "", "", ""
+    example = example_data[index_ex]
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return tasks, gaps, keywords, recent_works
+
+
+# Step 2: Generate Research Hypothesis and Experiment Plan
+def generate_and_store(paper_text, tasks, gaps, keywords, recent_works):
+    if (not state_extract or not state_example or paper_text == ""):
+        return "", "", "", ""
+    global state_generate, index_ex
+    state_generate = True
+    hypothesis = example_data[index_ex]['hypothesis']
+    experiment_plan = example_data[index_ex]['experiment_plan']
+    return hypothesis, experiment_plan, hypothesis, experiment_plan
+
+##########  Phase 2 & 3  ##############
+def start_experiment_agent(hypothesis, plan):
+    if (not state_extract or not state_generate or not state_example):
+        return "", "", ""
+    global state_agent, step_index, state_complete
+    state_agent = True
+    step_index = 0
+    state_complete = False
+    # predefined_message = f"Implement the following hypothesis and experiment plan:\n\nHypothesis:\n{hypothesis}\n\nExperiment Plan:\n{plan}"
+    return example_data[index_ex]['code_init'], predefined_action_log, "", ""
+
+def submit_feedback(user_feedback, history, previous_response):
+    if (not state_extract or not  state_generate or not state_agent or not state_example):
+        return "", "", ""
+    global step_index, state_complete
+    step_index += 1
+    msg = history
+    if step_index < len(process_steps):
+        msg += previous_response + "\nUser feedback:" + user_feedback + "\n\n"
+        response_info = process_steps[step_index]
+        response = info_to_message(response_info)  # Convert dictionary to formatted string
+        response += "Please provide feedback based on the history, response entries, and observation, and questions: "
+        step_index += 1
+        msg += response
+    else:
+        state_complete = True
+        response = "Agent Finished."
+        
+    return msg, response, example_data[index_ex]['code_init'] if state_complete else example_data[index_ex]['code_final'], ""
+
+def load_phase_2_inputs(hypothesis, plan):
+    return hypothesis, plan, "# Code implementation will be displayed here after Start ExperimentAgent."
+
+
+
+predefined_action_log = """
+[Reasoning]: To understand the initial structure and functionality of train.py for effective improvements.
+[Action]: Inspect Script (train.py)
+Input: {"script_name": "train.py", "start_line_number": "1", "end_line_number": "74"}
+Objective: Understand the training script, including data processing, [...]
+[Observation]: The train.py script imports [...]. Sets random seeds [...]. Defines [...] Placeholder functions [...] exist without implementation. [...]
+[Feedback]: The script structure is clear, but key functions (train_model, predict) need proper implementation for proposed model training and prediction.\n
+"""
+
+
+predefined_observation = """
+Epoch [1/10],
+Train MSE: 0.543,
+Test MSE: 0.688
+Epoch [2/10],
+Train MSE: 0.242,
+Test MSE: 0.493\n
+"""
+
+# Initialize the global step_index and history
+process_steps = [
+    {
+        "Action": "Inspect Script Lines (train.py)",
+        "Observation": (
+            "The train.py script imports necessary libraries (e.g., pandas, sklearn, torch). "
+            "Sets random seeds for reproducibility. Defines compute_metrics_for_regression function "
+            "to calculate RMSE for different dimensions. Placeholder functions train_model and "
+            "predict exist without implementations."
+        ),
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The script executed successfully. Generated embeddings using the BERT model. Completed "
+            "the training process without errors. Metrics calculation placeholders indicated areas needing implementation."
+        ),
+    },
+    {
+        "Action": "Edit Script (train.py)",
+        "Observation": (
+            "Edited train.py to separate data loading, model definition, training loop, and evaluation into distinct functions. "
+            "The edited train.py now has clearly defined functions"
+            "for data loading (load_data), model definition (build_model), "
+            "training (train_model), and evaluation (evaluate_model). Similarly, eval.py is reorganized to load the model and perform predictions efficiently."
+        ),
+    },
+    {
+        "Action": "Retrieve Model",
+        "Observation": "CNN and BiLSTM retrieved.",
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The model trained over the specified number of epochs. Training and validation loss values are recorded for each epoch, "
+            "the decrease in loss indicates improved model performance."
+        )
+    },
+    {
+        "Action": "Evaluation",
+        "Observation": predefined_observation,
+    }
+]
+def info_to_message(info):
+    msg = ""
+    for k, v in info.items():
+        if isinstance(v, dict):
+            tempv = v
+            v = ""
+            for k2, v2 in tempv.items():
+                v += f"{k2}:\n  {v2}\n"
+        v = User.indent_text(v, 2)
+        msg += '-' * 64
+        msg += '\n'
+        msg += f"{k}:\n{v}\n"
+    return msg
+
+
+def handle_example_click(example_index):
+    global index_ex
+    index_ex = example_index
+    return load_example(index_ex)  # Simply return the text to display it in the textbox
+
+# Gradio Interface
+with gr.Blocks(css=".gr-examples-label {display: none;}", theme=gr.themes.Default()) as app:
+    gr.Markdown("# MLR- Copilot: Machine Learning Research based on LLM Agents")
+    gr.Markdown("###   ")
+    gr.Markdown("## <span style='color:Orange;'> This UI is for predefined example demo only.</span>")
+    gr.Markdown("## <span style='color:Orange;'> To reproduce the results please use [Github](https://github.com/du-nlp-lab/MLR-Copilot/)</span>")
+
+
+
+    gr.Markdown("MLR-Copilot is a framework where LLMs mimic researchers’ thought processes, designed to enhance the productivity of machine learning research by automating the generation and implementation of research ideas. It begins with a research paper, autonomously generating and validating these ideas, while incorporating human feedback to help reach executable research outcomes.")
+
+
+    gr.Markdown("##    ⬇️  Click an example at bottom to start！")
+
+
+    # Use state variables to store generated hypothesis and experiment plan
+    hypothesis_state = gr.State("")
+    experiment_plan_state = gr.State("")
+
+    ########## Phase 1:  Research Idea Generation Tab ##############
+    with gr.Tab("💡Stage 1: Research Idea Generation"):
+        gr.Markdown("### Extract Research Elements and Generate Research Ideas")
+        with gr.Row():
+            with gr.Column():
+                paper_text_input = gr.Textbox(value="", lines=10, label="📑 Research Paper Text", interactive=False)
+
+                # extract_button = gr.Button("🔍 Extract Research Elements")
+                with gr.Row():
+                    tasks_output = gr.Textbox(placeholder="Research task definition", label="Research Tasks", lines=2, interactive=False)
+                    gaps_output = gr.Textbox(placeholder="Research gaps of current works", label="Research Gaps", lines=2, interactive=False)
+                    keywords_output = gr.Textbox(placeholder="Paper keywords", label="Keywords", lines=2, interactive=False)
+                    recent_works_output = gr.Textbox(placeholder="Recent works extracted from Semantic Scholar", label="Recent Works", lines=2, interactive=False)
+            with gr.Column():
+                with gr.Row():  # Move the button to the top 
+                    generate_button = gr.Button("✍️ Generate Research Hypothesis & Experiment Plan")
+                with gr.Group():
+                    gr.Markdown("### 🌟 Research Idea")
+                    with gr.Row():
+                        hypothesis_output = gr.Textbox(label="Generated Hypothesis", lines=20, interactive=False)
+                        experiment_plan_output = gr.Textbox(label="Generated Experiment Plan", lines=20, interactive=False)
+        
+    
+        # Step 1: Extract Research Elements
+        # extract_button.click(
+        #     fn=extract_research_elements, 
+        #     inputs=paper_text_input, 
+        #     outputs=[tasks_output, gaps_output, keywords_output, recent_works_output]
+        # )
+
+        generate_button.click(
+            fn=generate_and_store, 
+            inputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output], 
+            outputs=[hypothesis_output, experiment_plan_output, hypothesis_state, experiment_plan_state]
+        )
+
+        gr.Examples(
+            examples=example_text,
+            inputs=[paper_text_input],
+            outputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output, hypothesis_output, experiment_plan_output],
+            fn=load_example_and_set_index,
+            run_on_click = True,
+            # label="⬇️ Click an example to load"
+        )
+        
+
+
+
+    ########## Phase 2 & 3:  Experiment implementation and execution ##############
+    with gr.Tab("🧪 Stage 2 & Stage 3: Experiment implementation and execution"):
+        gr.Markdown("### Interact with the ExperimentAgent")
+
+        with gr.Row():
+            with gr.Column():
+                with gr.Group():
+                    gr.Markdown("### 🌟 Generated Research Idea")
+                    with gr.Row():
+                        idea_input = gr.Textbox(label="Generated Research Hypothesis", lines=30, interactive=False)
+                        plan_input = gr.Textbox(label="Generated Experiment Plan", lines=30, interactive=False)
+        
+            with gr.Column():
+                start_exp_agnet = gr.Button("⚙️ Start / Reset ExperimentAgent", elem_classes=["agent-btn"])
+                with gr.Group():
+                    gr.Markdown("### Implementation + Execution Log")    
+                    log = gr.Textbox(label="📖 Execution Log", lines=20, interactive=False)
+                    code_display = gr.Code(label="🧑‍💻 Implementation", language="python", interactive=False)
+
+            with gr.Column():
+                response = gr.Textbox(label="🤖 ExperimentAgent Response", lines=30, interactive=False)
+                feedback = gr.Textbox(placeholder="N/A", label="🧑‍🔬 User Feedback", lines=3, interactive=True)
+                submit_button = gr.Button("Submit", elem_classes=["Submit-btn"])
+        
+        hypothesis_state.change(
+            fn=load_phase_2_inputs, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[idea_input, plan_input, code_display]
+        )
+        
+        # Start research agent
+        start_exp_agnet.click(
+            fn=start_experiment_agent, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[code_display, log, response, feedback]
+        )
+        
+        submit_button.click(
+            fn=submit_feedback,
+            inputs=[feedback, log, response],
+            outputs=[log, response, code_display, feedback]
+        )
+
+# Test
+if __name__ == "__main__":
+    step_index = 0
+    app.launch(share=True)

.history/app_20250404174823.py

@@ -0,0 +1,335 @@
+import gradio as gr
+from pathlib import Path
+from reactagent.environment import Environment
+from reactagent.agents.agent_research import ResearchAgent
+from reactagent.runner import create_parser
+from reactagent import llm
+from reactagent.users.user import User
+import os
+import json
+
+
+# Global variables to store session state
+env = None
+agent = None
+state_example = False
+state_extract = False
+state_generate = False
+state_agent = False
+state_complete = False
+index_ex = "1"
+
+example_text = [
+    "Research Paper 1: Dataset and Baseline for Automatic Student Feedback Analysis",
+    "Research Paper 2: An Empirical Study on the Impact of Code Review on Software Quality"
+]
+
+
+# Load example JSON file
+def load_example_data():
+    with open("example/example_data.json", "r") as json_file:
+        example_data = json.load(json_file)
+    
+    for idx in example_data.keys():
+        try:
+            file = example_data[idx]["code_init"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_init"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")
+        try:
+            file = example_data[idx]["code_final"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_final"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")     
+    return example_data
+
+example_data = load_example_data()
+
+# Function to handle the selection of an example and populate the respective fields
+def load_example(example_id):
+    global index_ex 
+    index_ex = str(example_id)
+    example = example_data[index_ex]
+    paper_text = 'Title:\t' +  example['title'] + '\n\nAbstract:\t' + example['abstract']
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return [paper_text, tasks, gaps, keywords, recent_works]
+
+example_text = [load_example(1)[0], load_example(2)[0]]
+
+
+# Function to handle example clicks
+def load_example_and_set_index(paper_text_input):
+    global index_ex, state_example
+    state_example = True
+    index_ex = str(example_text.index(paper_text_input) + 1)
+    example =  load_example(index_ex)
+
+    return example
+
+
+
+########## Phase 1 ##############
+
+def extract_research_elements(paper_text):
+    global state_extract, index_ex, state_example
+    if not state_example or paper_text == "":
+        return "", "", "", ""
+    state_extract = True
+    if not paper_text.strip().startswith("Title:\t" + example_data[index_ex]["title"]):
+        print("Mismatch detected.")
+        print(paper_text)
+        return "", "", "", ""
+    example = example_data[index_ex]
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return tasks, gaps, keywords, recent_works
+
+
+# Step 2: Generate Research Hypothesis and Experiment Plan
+def generate_and_store(paper_text, tasks, gaps, keywords, recent_works):
+    if (not state_extract or not state_example or paper_text == ""):
+        return "", "", "", ""
+    global state_generate, index_ex
+    state_generate = True
+    hypothesis = example_data[index_ex]['hypothesis']
+    experiment_plan = example_data[index_ex]['experiment_plan']
+    return hypothesis, experiment_plan, hypothesis, experiment_plan
+
+##########  Phase 2 & 3  ##############
+def start_experiment_agent(hypothesis, plan):
+    if (not state_extract or not state_generate or not state_example):
+        return "", "", ""
+    global state_agent, step_index, state_complete
+    state_agent = True
+    step_index = 0
+    state_complete = False
+    # predefined_message = f"Implement the following hypothesis and experiment plan:\n\nHypothesis:\n{hypothesis}\n\nExperiment Plan:\n{plan}"
+    return example_data[index_ex]['code_init'], predefined_action_log, "", ""
+
+def submit_feedback(user_feedback, history, previous_response):
+    if (not state_extract or not  state_generate or not state_agent or not state_example):
+        return "", "", ""
+    global step_index, state_complete
+    step_index += 1
+    msg = history
+    if step_index < len(process_steps):
+        msg += previous_response + "\nUser feedback:" + user_feedback + "\n\n"
+        response_info = process_steps[step_index]
+        response = info_to_message(response_info)  # Convert dictionary to formatted string
+        response += "Please provide feedback based on the history, response entries, and observation, and questions: "
+        step_index += 1
+        msg += response
+    else:
+        state_complete = True
+        response = "Agent Finished."
+        
+    return msg, response, example_data[index_ex]['code_init'] if state_complete else example_data[index_ex]['code_final'], ""
+
+def load_phase_2_inputs(hypothesis, plan):
+    return hypothesis, plan, "# Code implementation will be displayed here after Start ExperimentAgent."
+
+
+
+predefined_action_log = """
+[Reasoning]: To understand the initial structure and functionality of train.py for effective improvements.
+[Action]: Inspect Script (train.py)
+Input: {"script_name": "train.py", "start_line_number": "1", "end_line_number": "74"}
+Objective: Understand the training script, including data processing, [...]
+[Observation]: The train.py script imports [...]. Sets random seeds [...]. Defines [...] Placeholder functions [...] exist without implementation. [...]
+[Feedback]: The script structure is clear, but key functions (train_model, predict) need proper implementation for proposed model training and prediction.\n
+"""
+
+
+predefined_observation = """
+Epoch [1/10],
+Train MSE: 0.543,
+Test MSE: 0.688
+Epoch [2/10],
+Train MSE: 0.242,
+Test MSE: 0.493\n
+"""
+
+# Initialize the global step_index and history
+process_steps = [
+    {
+        "Action": "Inspect Script Lines (train.py)",
+        "Observation": (
+            "The train.py script imports necessary libraries (e.g., pandas, sklearn, torch). "
+            "Sets random seeds for reproducibility. Defines compute_metrics_for_regression function "
+            "to calculate RMSE for different dimensions. Placeholder functions train_model and "
+            "predict exist without implementations."
+        ),
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The script executed successfully. Generated embeddings using the BERT model. Completed "
+            "the training process without errors. Metrics calculation placeholders indicated areas needing implementation."
+        ),
+    },
+    {
+        "Action": "Edit Script (train.py)",
+        "Observation": (
+            "Edited train.py to separate data loading, model definition, training loop, and evaluation into distinct functions. "
+            "The edited train.py now has clearly defined functions"
+            "for data loading (load_data), model definition (build_model), "
+            "training (train_model), and evaluation (evaluate_model). Similarly, eval.py is reorganized to load the model and perform predictions efficiently."
+        ),
+    },
+    {
+        "Action": "Retrieve Model",
+        "Observation": "CNN and BiLSTM retrieved.",
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The model trained over the specified number of epochs. Training and validation loss values are recorded for each epoch, "
+            "the decrease in loss indicates improved model performance."
+        )
+    },
+    {
+        "Action": "Evaluation",
+        "Observation": predefined_observation,
+    }
+]
+def info_to_message(info):
+    msg = ""
+    for k, v in info.items():
+        if isinstance(v, dict):
+            tempv = v
+            v = ""
+            for k2, v2 in tempv.items():
+                v += f"{k2}:\n  {v2}\n"
+        v = User.indent_text(v, 2)
+        msg += '-' * 64
+        msg += '\n'
+        msg += f"{k}:\n{v}\n"
+    return msg
+
+
+def handle_example_click(example_index):
+    global index_ex
+    index_ex = example_index
+    return load_example(index_ex)  # Simply return the text to display it in the textbox
+
+# Gradio Interface
+with gr.Blocks(css=".gr-examples-label {display: none;}", theme=gr.themes.Default()) as app:
+    gr.Markdown("# MLR- Copilot: Machine Learning Research based on LLM Agents")
+    gr.Markdown("###   ")
+    gr.Markdown("## <span style='color:Orange;'> This UI is for predefined example demo only.</span>")
+    gr.Markdown("## <span style='color:Orange;'> To reproduce the results please use [Github](https://github.com/du-nlp-lab/MLR-Copilot/)</span>")
+
+
+
+    gr.Markdown("MLR-Copilot is a framework where LLMs mimic researchers’ thought processes, designed to enhance the productivity of machine learning research by automating the generation and implementation of research ideas. It begins with a research paper, autonomously generating and validating these ideas, while incorporating human feedback to help reach executable research outcomes.")
+
+
+    gr.Markdown("##    ⬇️  Click an example at bottom to start！")
+
+
+    # Use state variables to store generated hypothesis and experiment plan
+    hypothesis_state = gr.State("")
+    experiment_plan_state = gr.State("")
+
+    ########## Phase 1:  Research Idea Generation Tab ##############
+    with gr.Tab("💡Stage 1: Research Idea Generation"):
+        gr.Markdown("### Extract Research Elements and Generate Research Ideas")
+        with gr.Row():
+            with gr.Column():
+                paper_text_input = gr.Textbox(value="", lines=10, label="📑 Research Paper Text", interactive=False)
+
+                # extract_button = gr.Button("🔍 Extract Research Elements")
+                with gr.Row():
+                    tasks_output = gr.Textbox(placeholder="Research task definition", label="Research Tasks", lines=2, interactive=False)
+                    gaps_output = gr.Textbox(placeholder="Research gaps of current works", label="Research Gaps", lines=2, interactive=False)
+                    keywords_output = gr.Textbox(placeholder="Paper keywords", label="Keywords", lines=2, interactive=False)
+                    recent_works_output = gr.Textbox(placeholder="Recent works extracted from Semantic Scholar", label="Recent Works", lines=2, interactive=False)
+            with gr.Column():
+                with gr.Row():  # Move the button to the top 
+                    generate_button = gr.Button("✍️ Generate Research Hypothesis & Experiment Plan")
+                with gr.Group():
+                    gr.Markdown("### 🌟 Research Idea")
+                    with gr.Row():
+                        hypothesis_output = gr.Textbox(label="Generated Hypothesis", lines=20, interactive=False)
+                        experiment_plan_output = gr.Textbox(label="Generated Experiment Plan", lines=20, interactive=False)
+        
+    
+        # Step 1: Extract Research Elements
+        # extract_button.click(
+        #     fn=extract_research_elements, 
+        #     inputs=paper_text_input, 
+        #     outputs=[tasks_output, gaps_output, keywords_output, recent_works_output]
+        # )
+
+        generate_button.click(
+            fn=generate_and_store, 
+            inputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output], 
+            outputs=[hypothesis_output, experiment_plan_output, hypothesis_state, experiment_plan_state]
+        )
+
+        gr.Examples(
+            examples=example_text,
+            inputs=[paper_text_input],
+            outputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output],
+            fn=load_example_and_set_index,
+            run_on_click = True,
+            # label="⬇️ Click an example to load"
+        )
+        
+
+
+
+    ########## Phase 2 & 3:  Experiment implementation and execution ##############
+    with gr.Tab("🧪 Stage 2 & Stage 3: Experiment implementation and execution"):
+        gr.Markdown("### Interact with the ExperimentAgent")
+
+        with gr.Row():
+            with gr.Column():
+                with gr.Group():
+                    gr.Markdown("### 🌟 Generated Research Idea")
+                    with gr.Row():
+                        idea_input = gr.Textbox(label="Generated Research Hypothesis", lines=30, interactive=False)
+                        plan_input = gr.Textbox(label="Generated Experiment Plan", lines=30, interactive=False)
+        
+            with gr.Column():
+                start_exp_agnet = gr.Button("⚙️ Start / Reset ExperimentAgent", elem_classes=["agent-btn"])
+                with gr.Group():
+                    gr.Markdown("### Implementation + Execution Log")    
+                    log = gr.Textbox(label="📖 Execution Log", lines=20, interactive=False)
+                    code_display = gr.Code(label="🧑‍💻 Implementation", language="python", interactive=False)
+
+            with gr.Column():
+                response = gr.Textbox(label="🤖 ExperimentAgent Response", lines=30, interactive=False)
+                feedback = gr.Textbox(placeholder="N/A", label="🧑‍🔬 User Feedback", lines=3, interactive=True)
+                submit_button = gr.Button("Submit", elem_classes=["Submit-btn"])
+        
+        hypothesis_state.change(
+            fn=load_phase_2_inputs, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[idea_input, plan_input, code_display]
+        )
+        
+        # Start research agent
+        start_exp_agnet.click(
+            fn=start_experiment_agent, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[code_display, log, response, feedback]
+        )
+        
+        submit_button.click(
+            fn=submit_feedback,
+            inputs=[feedback, log, response],
+            outputs=[log, response, code_display, feedback]
+        )
+
+# Test
+if __name__ == "__main__":
+    step_index = 0
+    app.launch(share=True)

.history/app_20250404174828.py

@@ -0,0 +1,335 @@
+import gradio as gr
+from pathlib import Path
+from reactagent.environment import Environment
+from reactagent.agents.agent_research import ResearchAgent
+from reactagent.runner import create_parser
+from reactagent import llm
+from reactagent.users.user import User
+import os
+import json
+
+
+# Global variables to store session state
+env = None
+agent = None
+state_example = False
+state_extract = False
+state_generate = False
+state_agent = False
+state_complete = False
+index_ex = "1"
+
+example_text = [
+    "Research Paper 1: Dataset and Baseline for Automatic Student Feedback Analysis",
+    "Research Paper 2: An Empirical Study on the Impact of Code Review on Software Quality"
+]
+
+
+# Load example JSON file
+def load_example_data():
+    with open("example/example_data.json", "r") as json_file:
+        example_data = json.load(json_file)
+    
+    for idx in example_data.keys():
+        try:
+            file = example_data[idx]["code_init"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_init"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")
+        try:
+            file = example_data[idx]["code_final"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_final"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")     
+    return example_data
+
+example_data = load_example_data()
+
+# Function to handle the selection of an example and populate the respective fields
+def load_example(example_id):
+    global index_ex 
+    index_ex = str(example_id)
+    example = example_data[index_ex]
+    paper_text = 'Title:\t' +  example['title'] + '\n\nAbstract:\t' + example['abstract']
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return [paper_text, tasks, gaps, keywords, recent_works]
+
+example_text = [load_example(1)[0], load_example(2)[0]]
+
+
+# Function to handle example clicks
+def load_example_and_set_index(paper_text_input):
+    global index_ex, state_example
+    state_example = True
+    index_ex = str(example_text.index(paper_text_input) + 1)
+    example =  load_example(index_ex)
+
+    return example
+
+
+
+########## Phase 1 ##############
+
+def extract_research_elements(paper_text):
+    global state_extract, index_ex, state_example
+    if not state_example or paper_text == "":
+        return "", "", "", ""
+    state_extract = True
+    if not paper_text.strip().startswith("Title:\t" + example_data[index_ex]["title"]):
+        print("Mismatch detected.")
+        print(paper_text)
+        return "", "", "", ""
+    example = example_data[index_ex]
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return tasks, gaps, keywords, recent_works
+
+
+# Step 2: Generate Research Hypothesis and Experiment Plan
+def generate_and_store(paper_text, tasks, gaps, keywords, recent_works):
+    if (not state_extract or not state_example or paper_text == ""):
+        return "", "", "", ""
+    global state_generate, index_ex
+    state_generate = True
+    hypothesis = example_data[index_ex]['hypothesis']
+    experiment_plan = example_data[index_ex]['experiment_plan']
+    return hypothesis, experiment_plan, hypothesis, experiment_plan
+
+##########  Phase 2 & 3  ##############
+def start_experiment_agent(hypothesis, plan):
+    if (not state_extract or not state_generate or not state_example):
+        return "", "", ""
+    global state_agent, step_index, state_complete
+    state_agent = True
+    step_index = 0
+    state_complete = False
+    # predefined_message = f"Implement the following hypothesis and experiment plan:\n\nHypothesis:\n{hypothesis}\n\nExperiment Plan:\n{plan}"
+    return example_data[index_ex]['code_init'], predefined_action_log, "", ""
+
+def submit_feedback(user_feedback, history, previous_response):
+    if (not state_extract or not  state_generate or not state_agent or not state_example):
+        return "", "", ""
+    global step_index, state_complete
+    step_index += 1
+    msg = history
+    if step_index < len(process_steps):
+        msg += previous_response + "\nUser feedback:" + user_feedback + "\n\n"
+        response_info = process_steps[step_index]
+        response = info_to_message(response_info)  # Convert dictionary to formatted string
+        response += "Please provide feedback based on the history, response entries, and observation, and questions: "
+        step_index += 1
+        msg += response
+    else:
+        state_complete = True
+        response = "Agent Finished."
+        
+    return msg, response, example_data[index_ex]['code_init'] if state_complete else example_data[index_ex]['code_final'], ""
+
+def load_phase_2_inputs(hypothesis, plan):
+    return hypothesis, plan, "# Code implementation will be displayed here after Start ExperimentAgent."
+
+
+
+predefined_action_log = """
+[Reasoning]: To understand the initial structure and functionality of train.py for effective improvements.
+[Action]: Inspect Script (train.py)
+Input: {"script_name": "train.py", "start_line_number": "1", "end_line_number": "74"}
+Objective: Understand the training script, including data processing, [...]
+[Observation]: The train.py script imports [...]. Sets random seeds [...]. Defines [...] Placeholder functions [...] exist without implementation. [...]
+[Feedback]: The script structure is clear, but key functions (train_model, predict) need proper implementation for proposed model training and prediction.\n
+"""
+
+
+predefined_observation = """
+Epoch [1/10],
+Train MSE: 0.543,
+Test MSE: 0.688
+Epoch [2/10],
+Train MSE: 0.242,
+Test MSE: 0.493\n
+"""
+
+# Initialize the global step_index and history
+process_steps = [
+    {
+        "Action": "Inspect Script Lines (train.py)",
+        "Observation": (
+            "The train.py script imports necessary libraries (e.g., pandas, sklearn, torch). "
+            "Sets random seeds for reproducibility. Defines compute_metrics_for_regression function "
+            "to calculate RMSE for different dimensions. Placeholder functions train_model and "
+            "predict exist without implementations."
+        ),
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The script executed successfully. Generated embeddings using the BERT model. Completed "
+            "the training process without errors. Metrics calculation placeholders indicated areas needing implementation."
+        ),
+    },
+    {
+        "Action": "Edit Script (train.py)",
+        "Observation": (
+            "Edited train.py to separate data loading, model definition, training loop, and evaluation into distinct functions. "
+            "The edited train.py now has clearly defined functions"
+            "for data loading (load_data), model definition (build_model), "
+            "training (train_model), and evaluation (evaluate_model). Similarly, eval.py is reorganized to load the model and perform predictions efficiently."
+        ),
+    },
+    {
+        "Action": "Retrieve Model",
+        "Observation": "CNN and BiLSTM retrieved.",
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The model trained over the specified number of epochs. Training and validation loss values are recorded for each epoch, "
+            "the decrease in loss indicates improved model performance."
+        )
+    },
+    {
+        "Action": "Evaluation",
+        "Observation": predefined_observation,
+    }
+]
+def info_to_message(info):
+    msg = ""
+    for k, v in info.items():
+        if isinstance(v, dict):
+            tempv = v
+            v = ""
+            for k2, v2 in tempv.items():
+                v += f"{k2}:\n  {v2}\n"
+        v = User.indent_text(v, 2)
+        msg += '-' * 64
+        msg += '\n'
+        msg += f"{k}:\n{v}\n"
+    return msg
+
+
+def handle_example_click(example_index):
+    global index_ex
+    index_ex = example_index
+    return load_example(index_ex)  # Simply return the text to display it in the textbox
+
+# Gradio Interface
+with gr.Blocks(css=".gr-examples-label {display: none;}", theme=gr.themes.Default()) as app:
+    gr.Markdown("# MLR- Copilot: Machine Learning Research based on LLM Agents")
+    gr.Markdown("###   ")
+    gr.Markdown("## <span style='color:Orange;'> This UI is for predefined example demo only.</span>")
+    gr.Markdown("## <span style='color:Orange;'> To reproduce the results please use [Github](https://github.com/du-nlp-lab/MLR-Copilot/)</span>")
+
+
+
+    gr.Markdown("MLR-Copilot is a framework where LLMs mimic researchers’ thought processes, designed to enhance the productivity of machine learning research by automating the generation and implementation of research ideas. It begins with a research paper, autonomously generating and validating these ideas, while incorporating human feedback to help reach executable research outcomes.")
+
+
+    gr.Markdown("##    ⬇️  Click an example at bottom to start！")
+
+
+    # Use state variables to store generated hypothesis and experiment plan
+    hypothesis_state = gr.State("")
+    experiment_plan_state = gr.State("")
+
+    ########## Phase 1:  Research Idea Generation Tab ##############
+    with gr.Tab("💡Stage 1: Research Idea Generation"):
+        gr.Markdown("### Extract Research Elements and Generate Research Ideas")
+        with gr.Row():
+            with gr.Column():
+                paper_text_input = gr.Textbox(value="", lines=10, label="📑 Research Paper Text", interactive=False)
+
+                # extract_button = gr.Button("🔍 Extract Research Elements")
+                with gr.Row():
+                    tasks_output = gr.Textbox(placeholder="Research task definition", label="Research Tasks", lines=2, interactive=False)
+                    gaps_output = gr.Textbox(placeholder="Research gaps of current works", label="Research Gaps", lines=2, interactive=False)
+                    keywords_output = gr.Textbox(placeholder="Paper keywords", label="Keywords", lines=2, interactive=False)
+                    recent_works_output = gr.Textbox(placeholder="Recent works extracted from Semantic Scholar", label="Recent Works", lines=2, interactive=False)
+            with gr.Column():
+                with gr.Row():  # Move the button to the top 
+                    generate_button = gr.Button("✍️ Generate Research Hypothesis & Experiment Plan")
+                with gr.Group():
+                    gr.Markdown("### 🌟 Research Idea")
+                    with gr.Row():
+                        hypothesis_output = gr.Textbox(label="Generated Hypothesis", lines=20, interactive=False)
+                        experiment_plan_output = gr.Textbox(label="Generated Experiment Plan", lines=20, interactive=False)
+        
+    
+        # Step 1: Extract Research Elements
+        # extract_button.click(
+        #     fn=extract_research_elements, 
+        #     inputs=paper_text_input, 
+        #     outputs=[tasks_output, gaps_output, keywords_output, recent_works_output]
+        # )
+
+        generate_button.click(
+            fn=generate_and_store, 
+            inputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output], 
+            outputs=[hypothesis_output, experiment_plan_output, hypothesis_state, experiment_plan_state]
+        )
+
+        gr.Examples(
+            examples=example_text,
+            inputs=[paper_text_input],
+            outputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output],
+            fn=load_example_and_set_index,
+            run_on_click = True,
+            # label="⬇️ Click an example to load"
+        )
+        
+
+
+
+    ########## Phase 2 & 3:  Experiment implementation and execution ##############
+    with gr.Tab("🧪 Stage 2 & Stage 3: Experiment implementation and execution"):
+        gr.Markdown("### Interact with the ExperimentAgent")
+
+        with gr.Row():
+            with gr.Column():
+                with gr.Group():
+                    gr.Markdown("### 🌟 Generated Research Idea")
+                    with gr.Row():
+                        idea_input = gr.Textbox(label="Generated Research Hypothesis", lines=30, interactive=False)
+                        plan_input = gr.Textbox(label="Generated Experiment Plan", lines=30, interactive=False)
+        
+            with gr.Column():
+                start_exp_agnet = gr.Button("⚙️ Start / Reset ExperimentAgent", elem_classes=["agent-btn"])
+                with gr.Group():
+                    gr.Markdown("### Implementation + Execution Log")    
+                    log = gr.Textbox(label="📖 Execution Log", lines=20, interactive=False)
+                    code_display = gr.Code(label="🧑‍💻 Implementation", language="python", interactive=False)
+
+            with gr.Column():
+                response = gr.Textbox(label="🤖 ExperimentAgent Response", lines=30, interactive=False)
+                feedback = gr.Textbox(placeholder="N/A", label="🧑‍🔬 User Feedback", lines=3, interactive=True)
+                submit_button = gr.Button("Submit", elem_classes=["Submit-btn"])
+        
+        hypothesis_state.change(
+            fn=load_phase_2_inputs, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[idea_input, plan_input, code_display]
+        )
+        
+        # Start research agent
+        start_exp_agnet.click(
+            fn=start_experiment_agent, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[code_display, log, response, feedback]
+        )
+        
+        submit_button.click(
+            fn=submit_feedback,
+            inputs=[feedback, log, response],
+            outputs=[log, response, code_display, feedback]
+        )
+
+# Test
+if __name__ == "__main__":
+    step_index = 0
+    app.launch(share=True)

.history/app_20250404174831.py

@@ -0,0 +1,335 @@
+import gradio as gr
+from pathlib import Path
+from reactagent.environment import Environment
+from reactagent.agents.agent_research import ResearchAgent
+from reactagent.runner import create_parser
+from reactagent import llm
+from reactagent.users.user import User
+import os
+import json
+
+
+# Global variables to store session state
+env = None
+agent = None
+state_example = False
+state_extract = False
+state_generate = False
+state_agent = False
+state_complete = False
+index_ex = "1"
+
+example_text = [
+    "Research Paper 1: Dataset and Baseline for Automatic Student Feedback Analysis",
+    "Research Paper 2: An Empirical Study on the Impact of Code Review on Software Quality"
+]
+
+
+# Load example JSON file
+def load_example_data():
+    with open("example/example_data.json", "r") as json_file:
+        example_data = json.load(json_file)
+    
+    for idx in example_data.keys():
+        try:
+            file = example_data[idx]["code_init"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_init"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")
+        try:
+            file = example_data[idx]["code_final"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_final"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")     
+    return example_data
+
+example_data = load_example_data()
+
+# Function to handle the selection of an example and populate the respective fields
+def load_example(example_id):
+    global index_ex 
+    index_ex = str(example_id)
+    example = example_data[index_ex]
+    paper_text = 'Title:\t' +  example['title'] + '\n\nAbstract:\t' + example['abstract']
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return [paper_text, tasks, gaps, keywords, recent_works]
+
+example_text = [load_example(1)[0], load_example(2)[0]]
+
+
+# Function to handle example clicks
+def load_example_and_set_index(paper_text_input):
+    global index_ex, state_example
+    state_example = True
+    index_ex = str(example_text.index(paper_text_input) + 1)
+    example =  load_example(index_ex)
+
+    return example
+
+
+
+########## Phase 1 ##############
+
+def extract_research_elements(paper_text):
+    global state_extract, index_ex, state_example
+    if not state_example or paper_text == "":
+        return "", "", "", ""
+    state_extract = True
+    if not paper_text.strip().startswith("Title:\t" + example_data[index_ex]["title"]):
+        print("Mismatch detected.")
+        print(paper_text)
+        return "", "", "", ""
+    example = example_data[index_ex]
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return tasks, gaps, keywords, recent_works
+
+
+# Step 2: Generate Research Hypothesis and Experiment Plan
+def generate_and_store(paper_text, tasks, gaps, keywords, recent_works):
+    if (not state_extract or not state_example or paper_text == ""):
+        return "", "", "", ""
+    global state_generate, index_ex
+    state_generate = True
+    hypothesis = example_data[index_ex]['hypothesis']
+    experiment_plan = example_data[index_ex]['experiment_plan']
+    return hypothesis, experiment_plan, hypothesis, experiment_plan
+
+##########  Phase 2 & 3  ##############
+def start_experiment_agent(hypothesis, plan):
+    if (not state_extract or not state_generate or not state_example):
+        return "", "", ""
+    global state_agent, step_index, state_complete
+    state_agent = True
+    step_index = 0
+    state_complete = False
+    # predefined_message = f"Implement the following hypothesis and experiment plan:\n\nHypothesis:\n{hypothesis}\n\nExperiment Plan:\n{plan}"
+    return example_data[index_ex]['code_init'], predefined_action_log, "", ""
+
+def submit_feedback(user_feedback, history, previous_response):
+    if (not state_extract or not  state_generate or not state_agent or not state_example):
+        return "", "", ""
+    global step_index, state_complete
+    step_index += 1
+    msg = history
+    if step_index < len(process_steps):
+        msg += previous_response + "\nUser feedback:" + user_feedback + "\n\n"
+        response_info = process_steps[step_index]
+        response = info_to_message(response_info)  # Convert dictionary to formatted string
+        response += "Please provide feedback based on the history, response entries, and observation, and questions: "
+        step_index += 1
+        msg += response
+    else:
+        state_complete = True
+        response = "Agent Finished."
+        
+    return msg, response, example_data[index_ex]['code_init'] if state_complete else example_data[index_ex]['code_final'], ""
+
+def load_phase_2_inputs(hypothesis, plan):
+    return hypothesis, plan, "# Code implementation will be displayed here after Start ExperimentAgent."
+
+
+
+predefined_action_log = """
+[Reasoning]: To understand the initial structure and functionality of train.py for effective improvements.
+[Action]: Inspect Script (train.py)
+Input: {"script_name": "train.py", "start_line_number": "1", "end_line_number": "74"}
+Objective: Understand the training script, including data processing, [...]
+[Observation]: The train.py script imports [...]. Sets random seeds [...]. Defines [...] Placeholder functions [...] exist without implementation. [...]
+[Feedback]: The script structure is clear, but key functions (train_model, predict) need proper implementation for proposed model training and prediction.\n
+"""
+
+
+predefined_observation = """
+Epoch [1/10],
+Train MSE: 0.543,
+Test MSE: 0.688
+Epoch [2/10],
+Train MSE: 0.242,
+Test MSE: 0.493\n
+"""
+
+# Initialize the global step_index and history
+process_steps = [
+    {
+        "Action": "Inspect Script Lines (train.py)",
+        "Observation": (
+            "The train.py script imports necessary libraries (e.g., pandas, sklearn, torch). "
+            "Sets random seeds for reproducibility. Defines compute_metrics_for_regression function "
+            "to calculate RMSE for different dimensions. Placeholder functions train_model and "
+            "predict exist without implementations."
+        ),
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The script executed successfully. Generated embeddings using the BERT model. Completed "
+            "the training process without errors. Metrics calculation placeholders indicated areas needing implementation."
+        ),
+    },
+    {
+        "Action": "Edit Script (train.py)",
+        "Observation": (
+            "Edited train.py to separate data loading, model definition, training loop, and evaluation into distinct functions. "
+            "The edited train.py now has clearly defined functions"
+            "for data loading (load_data), model definition (build_model), "
+            "training (train_model), and evaluation (evaluate_model). Similarly, eval.py is reorganized to load the model and perform predictions efficiently."
+        ),
+    },
+    {
+        "Action": "Retrieve Model",
+        "Observation": "CNN and BiLSTM retrieved.",
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The model trained over the specified number of epochs. Training and validation loss values are recorded for each epoch, "
+            "the decrease in loss indicates improved model performance."
+        )
+    },
+    {
+        "Action": "Evaluation",
+        "Observation": predefined_observation,
+    }
+]
+def info_to_message(info):
+    msg = ""
+    for k, v in info.items():
+        if isinstance(v, dict):
+            tempv = v
+            v = ""
+            for k2, v2 in tempv.items():
+                v += f"{k2}:\n  {v2}\n"
+        v = User.indent_text(v, 2)
+        msg += '-' * 64
+        msg += '\n'
+        msg += f"{k}:\n{v}\n"
+    return msg
+
+
+def handle_example_click(example_index):
+    global index_ex
+    index_ex = example_index
+    return load_example(index_ex)  # Simply return the text to display it in the textbox
+
+# Gradio Interface
+with gr.Blocks(css=".gr-examples-label {display: none;}", theme=gr.themes.Default()) as app:
+    gr.Markdown("# MLR- Copilot: Machine Learning Research based on LLM Agents")
+    gr.Markdown("###   ")
+    gr.Markdown("## <span style='color:Orange;'> This UI is for predefined example demo only.</span>")
+    gr.Markdown("## <span style='color:Orange;'> To reproduce the results please use [Github](https://github.com/du-nlp-lab/MLR-Copilot/)</span>")
+
+
+
+    gr.Markdown("MLR-Copilot is a framework where LLMs mimic researchers’ thought processes, designed to enhance the productivity of machine learning research by automating the generation and implementation of research ideas. It begins with a research paper, autonomously generating and validating these ideas, while incorporating human feedback to help reach executable research outcomes.")
+
+
+    gr.Markdown("##    ⬇️  Click an example at bottom to start！")
+
+
+    # Use state variables to store generated hypothesis and experiment plan
+    hypothesis_state = gr.State("")
+    experiment_plan_state = gr.State("")
+
+    ########## Phase 1:  Research Idea Generation Tab ##############
+    with gr.Tab("💡Stage 1: Research Idea Generation"):
+        gr.Markdown("### Extract Research Elements and Generate Research Ideas")
+        with gr.Row():
+            with gr.Column():
+                paper_text_input = gr.Textbox(value="", lines=10, label="📑 Research Paper Text", interactive=False)
+
+                # extract_button = gr.Button("🔍 Extract Research Elements")
+                with gr.Row():
+                    tasks_output = gr.Textbox(placeholder="Research task definition", label="Research Tasks", lines=2, interactive=False)
+                    gaps_output = gr.Textbox(placeholder="Research gaps of current works", label="Research Gaps", lines=2, interactive=False)
+                    keywords_output = gr.Textbox(placeholder="Paper keywords", label="Keywords", lines=2, interactive=False)
+                    recent_works_output = gr.Textbox(placeholder="Recent works extracted from Semantic Scholar", label="Recent Works", lines=2, interactive=False)
+            with gr.Column():
+                with gr.Row():  # Move the button to the top 
+                    generate_button = gr.Button("✍️ Generate Research Hypothesis & Experiment Plan")
+                with gr.Group():
+                    gr.Markdown("### 🌟 Research Idea")
+                    with gr.Row():
+                        hypothesis_output = gr.Textbox(label="Generated Hypothesis", lines=20, interactive=False)
+                        experiment_plan_output = gr.Textbox(label="Generated Experiment Plan", lines=20, interactive=False)
+        
+    
+        # Step 1: Extract Research Elements
+        # extract_button.click(
+        #     fn=extract_research_elements, 
+        #     inputs=paper_text_input, 
+        #     outputs=[tasks_output, gaps_output, keywords_output, recent_works_output]
+        # )
+
+        generate_button.click(
+            fn=generate_and_store, 
+            inputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output], 
+            outputs=[hypothesis_output, experiment_plan_output, hypothesis_state, experiment_plan_state]
+        )
+
+        gr.Examples(
+            examples=example_text,
+            inputs=[paper_text_input],
+            outputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output],
+            fn=load_example_and_set_index,
+            run_on_click = True,
+            # label="⬇️ Click an example to load"
+        )
+        
+
+
+
+    ########## Phase 2 & 3:  Experiment implementation and execution ##############
+    with gr.Tab("🧪 Stage 2 & Stage 3: Experiment implementation and execution"):
+        gr.Markdown("### Interact with the ExperimentAgent")
+
+        with gr.Row():
+            with gr.Column():
+                with gr.Group():
+                    gr.Markdown("### 🌟 Generated Research Idea")
+                    with gr.Row():
+                        idea_input = gr.Textbox(label="Generated Research Hypothesis", lines=30, interactive=False)
+                        plan_input = gr.Textbox(label="Generated Experiment Plan", lines=30, interactive=False)
+        
+            with gr.Column():
+                start_exp_agnet = gr.Button("⚙️ Start / Reset ExperimentAgent", elem_classes=["agent-btn"])
+                with gr.Group():
+                    gr.Markdown("### Implementation + Execution Log")    
+                    log = gr.Textbox(label="📖 Execution Log", lines=20, interactive=False)
+                    code_display = gr.Code(label="🧑‍💻 Implementation", language="python", interactive=False)
+
+            with gr.Column():
+                response = gr.Textbox(label="🤖 ExperimentAgent Response", lines=30, interactive=False)
+                feedback = gr.Textbox(placeholder="N/A", label="🧑‍🔬 User Feedback", lines=3, interactive=True)
+                submit_button = gr.Button("Submit", elem_classes=["Submit-btn"])
+        
+        hypothesis_state.change(
+            fn=load_phase_2_inputs, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[idea_input, plan_input, code_display]
+        )
+        
+        # Start research agent
+        start_exp_agnet.click(
+            fn=start_experiment_agent, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[code_display, log, response, feedback]
+        )
+        
+        submit_button.click(
+            fn=submit_feedback,
+            inputs=[feedback, log, response],
+            outputs=[log, response, code_display, feedback]
+        )
+
+# Test
+if __name__ == "__main__":
+    step_index = 0
+    app.launch(share=True)

.history/app_20250404174832.py

@@ -0,0 +1,335 @@
+import gradio as gr
+from pathlib import Path
+from reactagent.environment import Environment
+from reactagent.agents.agent_research import ResearchAgent
+from reactagent.runner import create_parser
+from reactagent import llm
+from reactagent.users.user import User
+import os
+import json
+
+
+# Global variables to store session state
+env = None
+agent = None
+state_example = False
+state_extract = False
+state_generate = False
+state_agent = False
+state_complete = False
+index_ex = "1"
+
+example_text = [
+    "Research Paper 1: Dataset and Baseline for Automatic Student Feedback Analysis",
+    "Research Paper 2: An Empirical Study on the Impact of Code Review on Software Quality"
+]
+
+
+# Load example JSON file
+def load_example_data():
+    with open("example/example_data.json", "r") as json_file:
+        example_data = json.load(json_file)
+    
+    for idx in example_data.keys():
+        try:
+            file = example_data[idx]["code_init"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_init"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")
+        try:
+            file = example_data[idx]["code_final"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_final"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")     
+    return example_data
+
+example_data = load_example_data()
+
+# Function to handle the selection of an example and populate the respective fields
+def load_example(example_id):
+    global index_ex 
+    index_ex = str(example_id)
+    example = example_data[index_ex]
+    paper_text = 'Title:\t' +  example['title'] + '\n\nAbstract:\t' + example['abstract']
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return [paper_text, tasks, gaps, keywords, recent_works]
+
+example_text = [load_example(1)[0], load_example(2)[0]]
+
+
+# Function to handle example clicks
+def load_example_and_set_index(paper_text_input):
+    global index_ex, state_example
+    state_example = True
+    index_ex = str(example_text.index(paper_text_input) + 1)
+    example =  load_example(index_ex)
+
+    return example
+
+
+
+########## Phase 1 ##############
+
+def extract_research_elements(paper_text):
+    global state_extract, index_ex, state_example
+    if not state_example or paper_text == "":
+        return "", "", "", ""
+    state_extract = True
+    if not paper_text.strip().startswith("Title:\t" + example_data[index_ex]["title"]):
+        print("Mismatch detected.")
+        print(paper_text)
+        return "", "", "", ""
+    example = example_data[index_ex]
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return tasks, gaps, keywords, recent_works
+
+
+# Step 2: Generate Research Hypothesis and Experiment Plan
+def generate_and_store(paper_text, tasks, gaps, keywords, recent_works):
+    if (not state_extract or not state_example or paper_text == ""):
+        return "", "", "", ""
+    global state_generate, index_ex
+    state_generate = True
+    hypothesis = example_data[index_ex]['hypothesis']
+    experiment_plan = example_data[index_ex]['experiment_plan']
+    return hypothesis, experiment_plan, hypothesis, experiment_plan
+
+##########  Phase 2 & 3  ##############
+def start_experiment_agent(hypothesis, plan):
+    if (not state_extract or not state_generate or not state_example):
+        return "", "", ""
+    global state_agent, step_index, state_complete
+    state_agent = True
+    step_index = 0
+    state_complete = False
+    # predefined_message = f"Implement the following hypothesis and experiment plan:\n\nHypothesis:\n{hypothesis}\n\nExperiment Plan:\n{plan}"
+    return example_data[index_ex]['code_init'], predefined_action_log, "", ""
+
+def submit_feedback(user_feedback, history, previous_response):
+    if (not state_extract or not  state_generate or not state_agent or not state_example):
+        return "", "", ""
+    global step_index, state_complete
+    step_index += 1
+    msg = history
+    if step_index < len(process_steps):
+        msg += previous_response + "\nUser feedback:" + user_feedback + "\n\n"
+        response_info = process_steps[step_index]
+        response = info_to_message(response_info)  # Convert dictionary to formatted string
+        response += "Please provide feedback based on the history, response entries, and observation, and questions: "
+        step_index += 1
+        msg += response
+    else:
+        state_complete = True
+        response = "Agent Finished."
+        
+    return msg, response, example_data[index_ex]['code_init'] if state_complete else example_data[index_ex]['code_final'], ""
+
+def load_phase_2_inputs(hypothesis, plan):
+    return hypothesis, plan, "# Code implementation will be displayed here after Start ExperimentAgent."
+
+
+
+predefined_action_log = """
+[Reasoning]: To understand the initial structure and functionality of train.py for effective improvements.
+[Action]: Inspect Script (train.py)
+Input: {"script_name": "train.py", "start_line_number": "1", "end_line_number": "74"}
+Objective: Understand the training script, including data processing, [...]
+[Observation]: The train.py script imports [...]. Sets random seeds [...]. Defines [...] Placeholder functions [...] exist without implementation. [...]
+[Feedback]: The script structure is clear, but key functions (train_model, predict) need proper implementation for proposed model training and prediction.\n
+"""
+
+
+predefined_observation = """
+Epoch [1/10],
+Train MSE: 0.543,
+Test MSE: 0.688
+Epoch [2/10],
+Train MSE: 0.242,
+Test MSE: 0.493\n
+"""
+
+# Initialize the global step_index and history
+process_steps = [
+    {
+        "Action": "Inspect Script Lines (train.py)",
+        "Observation": (
+            "The train.py script imports necessary libraries (e.g., pandas, sklearn, torch). "
+            "Sets random seeds for reproducibility. Defines compute_metrics_for_regression function "
+            "to calculate RMSE for different dimensions. Placeholder functions train_model and "
+            "predict exist without implementations."
+        ),
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The script executed successfully. Generated embeddings using the BERT model. Completed "
+            "the training process without errors. Metrics calculation placeholders indicated areas needing implementation."
+        ),
+    },
+    {
+        "Action": "Edit Script (train.py)",
+        "Observation": (
+            "Edited train.py to separate data loading, model definition, training loop, and evaluation into distinct functions. "
+            "The edited train.py now has clearly defined functions"
+            "for data loading (load_data), model definition (build_model), "
+            "training (train_model), and evaluation (evaluate_model). Similarly, eval.py is reorganized to load the model and perform predictions efficiently."
+        ),
+    },
+    {
+        "Action": "Retrieve Model",
+        "Observation": "CNN and BiLSTM retrieved.",
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The model trained over the specified number of epochs. Training and validation loss values are recorded for each epoch, "
+            "the decrease in loss indicates improved model performance."
+        )
+    },
+    {
+        "Action": "Evaluation",
+        "Observation": predefined_observation,
+    }
+]
+def info_to_message(info):
+    msg = ""
+    for k, v in info.items():
+        if isinstance(v, dict):
+            tempv = v
+            v = ""
+            for k2, v2 in tempv.items():
+                v += f"{k2}:\n  {v2}\n"
+        v = User.indent_text(v, 2)
+        msg += '-' * 64
+        msg += '\n'
+        msg += f"{k}:\n{v}\n"
+    return msg
+
+
+def handle_example_click(example_index):
+    global index_ex
+    index_ex = example_index
+    return load_example(index_ex)  # Simply return the text to display it in the textbox
+
+# Gradio Interface
+with gr.Blocks(css=".gr-examples-label {display: none;}", theme=gr.themes.Default()) as app:
+    gr.Markdown("# MLR- Copilot: Machine Learning Research based on LLM Agents")
+    gr.Markdown("###   ")
+    gr.Markdown("## <span style='color:Orange;'> This UI is for predefined example demo only.</span>")
+    gr.Markdown("## <span style='color:Orange;'> To reproduce the results please use [Github](https://github.com/du-nlp-lab/MLR-Copilot/)</span>")
+
+
+
+    gr.Markdown("MLR-Copilot is a framework where LLMs mimic researchers’ thought processes, designed to enhance the productivity of machine learning research by automating the generation and implementation of research ideas. It begins with a research paper, autonomously generating and validating these ideas, while incorporating human feedback to help reach executable research outcomes.")
+
+
+    gr.Markdown("##    ⬇️  Click an example at bottom to start！")
+
+
+    # Use state variables to store generated hypothesis and experiment plan
+    hypothesis_state = gr.State("")
+    experiment_plan_state = gr.State("")
+
+    ########## Phase 1:  Research Idea Generation Tab ##############
+    with gr.Tab("💡Stage 1: Research Idea Generation"):
+        gr.Markdown("### Extract Research Elements and Generate Research Ideas")
+        with gr.Row():
+            with gr.Column():
+                paper_text_input = gr.Textbox(value="", lines=10, label="📑 Research Paper Text", interactive=False)
+
+                # extract_button = gr.Button("🔍 Extract Research Elements")
+                with gr.Row():
+                    tasks_output = gr.Textbox(placeholder="Research task definition", label="Research Tasks", lines=2, interactive=False)
+                    gaps_output = gr.Textbox(placeholder="Research gaps of current works", label="Research Gaps", lines=2, interactive=False)
+                    keywords_output = gr.Textbox(placeholder="Paper keywords", label="Keywords", lines=2, interactive=False)
+                    recent_works_output = gr.Textbox(placeholder="Recent works extracted from Semantic Scholar", label="Recent Works", lines=2, interactive=False)
+            with gr.Column():
+                with gr.Row():  # Move the button to the top 
+                    generate_button = gr.Button("✍️ Generate Research Hypothesis & Experiment Plan")
+                with gr.Group():
+                    gr.Markdown("### 🌟 Research Idea")
+                    with gr.Row():
+                        hypothesis_output = gr.Textbox(label="Generated Hypothesis", lines=20, interactive=False)
+                        experiment_plan_output = gr.Textbox(label="Generated Experiment Plan", lines=20, interactive=False)
+        
+    
+        # Step 1: Extract Research Elements
+        # extract_button.click(
+        #     fn=extract_research_elements, 
+        #     inputs=paper_text_input, 
+        #     outputs=[tasks_output, gaps_output, keywords_output, recent_works_output]
+        # )
+
+        generate_button.click(
+            fn=generate_and_store, 
+            inputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output], 
+            outputs=[hypothesis_output, experiment_plan_output, hypothesis_state, experiment_plan_state]
+        )
+
+        gr.Examples(
+            examples=example_text,
+            inputs=[paper_text_input],
+            outputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output],
+            fn=load_example_and_set_index,
+            run_on_click = True,
+            # label="⬇️ Click an example to load"
+        )
+        
+
+
+
+    ########## Phase 2 & 3:  Experiment implementation and execution ##############
+    with gr.Tab("🧪 Stage 2 & Stage 3: Experiment implementation and execution"):
+        gr.Markdown("### Interact with the ExperimentAgent")
+
+        with gr.Row():
+            with gr.Column():
+                with gr.Group():
+                    gr.Markdown("### 🌟 Generated Research Idea")
+                    with gr.Row():
+                        idea_input = gr.Textbox(label="Generated Research Hypothesis", lines=30, interactive=False)
+                        plan_input = gr.Textbox(label="Generated Experiment Plan", lines=30, interactive=False)
+        
+            with gr.Column():
+                start_exp_agnet = gr.Button("⚙️ Start / Reset ExperimentAgent", elem_classes=["agent-btn"])
+                with gr.Group():
+                    gr.Markdown("### Implementation + Execution Log")    
+                    log = gr.Textbox(label="📖 Execution Log", lines=20, interactive=False)
+                    code_display = gr.Code(label="🧑‍💻 Implementation", language="python", interactive=False)
+
+            with gr.Column():
+                response = gr.Textbox(label="🤖 ExperimentAgent Response", lines=30, interactive=False)
+                feedback = gr.Textbox(placeholder="N/A", label="🧑‍🔬 User Feedback", lines=3, interactive=True)
+                submit_button = gr.Button("Submit", elem_classes=["Submit-btn"])
+        
+        hypothesis_state.change(
+            fn=load_phase_2_inputs, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[idea_input, plan_input, code_display]
+        )
+        
+        # Start research agent
+        start_exp_agnet.click(
+            fn=start_experiment_agent, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[code_display, log, response, feedback]
+        )
+        
+        submit_button.click(
+            fn=submit_feedback,
+            inputs=[feedback, log, response],
+            outputs=[log, response, code_display, feedback]
+        )
+
+# Test
+if __name__ == "__main__":
+    step_index = 0
+    app.launch(share=True)

.history/app_20250404175024.py

@@ -0,0 +1,335 @@
+import gradio as gr
+from pathlib import Path
+from reactagent.environment import Environment
+from reactagent.agents.agent_research import ResearchAgent
+from reactagent.runner import create_parser
+from reactagent import llm
+from reactagent.users.user import User
+import os
+import json
+
+
+# Global variables to store session state
+env = None
+agent = None
+state_example = False
+state_extract = False
+state_generate = False
+state_agent = False
+state_complete = False
+index_ex = "1"
+
+example_text = [
+    "Research Paper 1: Dataset and Baseline for Automatic Student Feedback Analysis",
+    "Research Paper 2: An Empirical Study on the Impact of Code Review on Software Quality"
+]
+
+
+# Load example JSON file
+def load_example_data():
+    with open("example/example_data.json", "r") as json_file:
+        example_data = json.load(json_file)
+    
+    for idx in example_data.keys():
+        try:
+            file = example_data[idx]["code_init"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_init"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")
+        try:
+            file = example_data[idx]["code_final"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_final"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")     
+    return example_data
+
+example_data = load_example_data()
+
+# Function to handle the selection of an example and populate the respective fields
+def load_example(example_id):
+    global index_ex 
+    index_ex = str(example_id)
+    example = example_data[index_ex]
+    paper_text = 'Title:\t' +  example['title'] + '\n\nAbstract:\t' + example['abstract']
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return [paper_text, tasks, gaps, keywords, recent_works]
+
+example_text = [load_example(1)[0], load_example(2)[0]]
+
+
+# Function to handle example clicks
+def load_example_and_set_index(paper_text_input):
+    global index_ex, state_example
+    state_example = True
+    index_ex = str(example_text.index(paper_text_input) + 1)
+    example =  load_example(index_ex)
+
+    return example
+
+
+
+########## Phase 1 ##############
+
+def extract_research_elements(paper_text):
+    global state_extract, index_ex, state_example
+    if not state_example or paper_text == "":
+        return "", "", "", ""
+    state_extract = True
+    if not paper_text.strip().startswith("Title:\t" + example_data[index_ex]["title"]):
+        print("Mismatch detected.")
+        print(paper_text)
+        return "", "", "", ""
+    example = example_data[index_ex]
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return tasks, gaps, keywords, recent_works
+
+
+# Step 2: Generate Research Hypothesis and Experiment Plan
+def generate_and_store(paper_text, tasks, gaps, keywords, recent_works):
+    if (not state_extract or not state_example):
+        return "", "", "", ""
+    global state_generate, index_ex
+    state_generate = True
+    hypothesis = example_data[index_ex]['hypothesis']
+    experiment_plan = example_data[index_ex]['experiment_plan']
+    return hypothesis, experiment_plan, hypothesis, experiment_plan
+
+##########  Phase 2 & 3  ##############
+def start_experiment_agent(hypothesis, plan):
+    if (not state_extract or not state_generate or not state_example):
+        return "", "", ""
+    global state_agent, step_index, state_complete
+    state_agent = True
+    step_index = 0
+    state_complete = False
+    # predefined_message = f"Implement the following hypothesis and experiment plan:\n\nHypothesis:\n{hypothesis}\n\nExperiment Plan:\n{plan}"
+    return example_data[index_ex]['code_init'], predefined_action_log, "", ""
+
+def submit_feedback(user_feedback, history, previous_response):
+    if (not state_extract or not  state_generate or not state_agent or not state_example):
+        return "", "", ""
+    global step_index, state_complete
+    step_index += 1
+    msg = history
+    if step_index < len(process_steps):
+        msg += previous_response + "\nUser feedback:" + user_feedback + "\n\n"
+        response_info = process_steps[step_index]
+        response = info_to_message(response_info)  # Convert dictionary to formatted string
+        response += "Please provide feedback based on the history, response entries, and observation, and questions: "
+        step_index += 1
+        msg += response
+    else:
+        state_complete = True
+        response = "Agent Finished."
+        
+    return msg, response, example_data[index_ex]['code_init'] if state_complete else example_data[index_ex]['code_final'], ""
+
+def load_phase_2_inputs(hypothesis, plan):
+    return hypothesis, plan, "# Code implementation will be displayed here after Start ExperimentAgent."
+
+
+
+predefined_action_log = """
+[Reasoning]: To understand the initial structure and functionality of train.py for effective improvements.
+[Action]: Inspect Script (train.py)
+Input: {"script_name": "train.py", "start_line_number": "1", "end_line_number": "74"}
+Objective: Understand the training script, including data processing, [...]
+[Observation]: The train.py script imports [...]. Sets random seeds [...]. Defines [...] Placeholder functions [...] exist without implementation. [...]
+[Feedback]: The script structure is clear, but key functions (train_model, predict) need proper implementation for proposed model training and prediction.\n
+"""
+
+
+predefined_observation = """
+Epoch [1/10],
+Train MSE: 0.543,
+Test MSE: 0.688
+Epoch [2/10],
+Train MSE: 0.242,
+Test MSE: 0.493\n
+"""
+
+# Initialize the global step_index and history
+process_steps = [
+    {
+        "Action": "Inspect Script Lines (train.py)",
+        "Observation": (
+            "The train.py script imports necessary libraries (e.g., pandas, sklearn, torch). "
+            "Sets random seeds for reproducibility. Defines compute_metrics_for_regression function "
+            "to calculate RMSE for different dimensions. Placeholder functions train_model and "
+            "predict exist without implementations."
+        ),
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The script executed successfully. Generated embeddings using the BERT model. Completed "
+            "the training process without errors. Metrics calculation placeholders indicated areas needing implementation."
+        ),
+    },
+    {
+        "Action": "Edit Script (train.py)",
+        "Observation": (
+            "Edited train.py to separate data loading, model definition, training loop, and evaluation into distinct functions. "
+            "The edited train.py now has clearly defined functions"
+            "for data loading (load_data), model definition (build_model), "
+            "training (train_model), and evaluation (evaluate_model). Similarly, eval.py is reorganized to load the model and perform predictions efficiently."
+        ),
+    },
+    {
+        "Action": "Retrieve Model",
+        "Observation": "CNN and BiLSTM retrieved.",
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The model trained over the specified number of epochs. Training and validation loss values are recorded for each epoch, "
+            "the decrease in loss indicates improved model performance."
+        )
+    },
+    {
+        "Action": "Evaluation",
+        "Observation": predefined_observation,
+    }
+]
+def info_to_message(info):
+    msg = ""
+    for k, v in info.items():
+        if isinstance(v, dict):
+            tempv = v
+            v = ""
+            for k2, v2 in tempv.items():
+                v += f"{k2}:\n  {v2}\n"
+        v = User.indent_text(v, 2)
+        msg += '-' * 64
+        msg += '\n'
+        msg += f"{k}:\n{v}\n"
+    return msg
+
+
+def handle_example_click(example_index):
+    global index_ex
+    index_ex = example_index
+    return load_example(index_ex)  # Simply return the text to display it in the textbox
+
+# Gradio Interface
+with gr.Blocks(css=".gr-examples-label {display: none;}", theme=gr.themes.Default()) as app:
+    gr.Markdown("# MLR- Copilot: Machine Learning Research based on LLM Agents")
+    gr.Markdown("###   ")
+    gr.Markdown("## <span style='color:Orange;'> This UI is for predefined example demo only.</span>")
+    gr.Markdown("## <span style='color:Orange;'> To reproduce the results please use [Github](https://github.com/du-nlp-lab/MLR-Copilot/)</span>")
+
+
+
+    gr.Markdown("MLR-Copilot is a framework where LLMs mimic researchers’ thought processes, designed to enhance the productivity of machine learning research by automating the generation and implementation of research ideas. It begins with a research paper, autonomously generating and validating these ideas, while incorporating human feedback to help reach executable research outcomes.")
+
+
+    gr.Markdown("##    ⬇️  Click an example at bottom to start！")
+
+
+    # Use state variables to store generated hypothesis and experiment plan
+    hypothesis_state = gr.State("")
+    experiment_plan_state = gr.State("")
+
+    ########## Phase 1:  Research Idea Generation Tab ##############
+    with gr.Tab("💡Stage 1: Research Idea Generation"):
+        gr.Markdown("### Extract Research Elements and Generate Research Ideas")
+        with gr.Row():
+            with gr.Column():
+                paper_text_input = gr.Textbox(value="", lines=10, label="📑 Research Paper Text", interactive=False)
+
+                # extract_button = gr.Button("🔍 Extract Research Elements")
+                with gr.Row():
+                    tasks_output = gr.Textbox(placeholder="Research task definition", label="Research Tasks", lines=2, interactive=False)
+                    gaps_output = gr.Textbox(placeholder="Research gaps of current works", label="Research Gaps", lines=2, interactive=False)
+                    keywords_output = gr.Textbox(placeholder="Paper keywords", label="Keywords", lines=2, interactive=False)
+                    recent_works_output = gr.Textbox(placeholder="Recent works extracted from Semantic Scholar", label="Recent Works", lines=2, interactive=False)
+            with gr.Column():
+                with gr.Row():  # Move the button to the top 
+                    generate_button = gr.Button("✍️ Generate Research Hypothesis & Experiment Plan")
+                with gr.Group():
+                    gr.Markdown("### 🌟 Research Idea")
+                    with gr.Row():
+                        hypothesis_output = gr.Textbox(label="Generated Hypothesis", lines=20, interactive=False)
+                        experiment_plan_output = gr.Textbox(label="Generated Experiment Plan", lines=20, interactive=False)
+        
+    
+        # Step 1: Extract Research Elements
+        # extract_button.click(
+        #     fn=extract_research_elements, 
+        #     inputs=paper_text_input, 
+        #     outputs=[tasks_output, gaps_output, keywords_output, recent_works_output]
+        # )
+
+        generate_button.click(
+            fn=generate_and_store, 
+            inputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output], 
+            outputs=[hypothesis_output, experiment_plan_output, hypothesis_state, experiment_plan_state]
+        )
+
+        gr.Examples(
+            examples=example_text,
+            inputs=[paper_text_input],
+            outputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output],
+            fn=load_example_and_set_index,
+            run_on_click = True,
+            # label="⬇️ Click an example to load"
+        )
+        
+
+
+
+    ########## Phase 2 & 3:  Experiment implementation and execution ##############
+    with gr.Tab("🧪 Stage 2 & Stage 3: Experiment implementation and execution"):
+        gr.Markdown("### Interact with the ExperimentAgent")
+
+        with gr.Row():
+            with gr.Column():
+                with gr.Group():
+                    gr.Markdown("### 🌟 Generated Research Idea")
+                    with gr.Row():
+                        idea_input = gr.Textbox(label="Generated Research Hypothesis", lines=30, interactive=False)
+                        plan_input = gr.Textbox(label="Generated Experiment Plan", lines=30, interactive=False)
+        
+            with gr.Column():
+                start_exp_agnet = gr.Button("⚙️ Start / Reset ExperimentAgent", elem_classes=["agent-btn"])
+                with gr.Group():
+                    gr.Markdown("### Implementation + Execution Log")    
+                    log = gr.Textbox(label="📖 Execution Log", lines=20, interactive=False)
+                    code_display = gr.Code(label="🧑‍💻 Implementation", language="python", interactive=False)
+
+            with gr.Column():
+                response = gr.Textbox(label="🤖 ExperimentAgent Response", lines=30, interactive=False)
+                feedback = gr.Textbox(placeholder="N/A", label="🧑‍🔬 User Feedback", lines=3, interactive=True)
+                submit_button = gr.Button("Submit", elem_classes=["Submit-btn"])
+        
+        hypothesis_state.change(
+            fn=load_phase_2_inputs, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[idea_input, plan_input, code_display]
+        )
+        
+        # Start research agent
+        start_exp_agnet.click(
+            fn=start_experiment_agent, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[code_display, log, response, feedback]
+        )
+        
+        submit_button.click(
+            fn=submit_feedback,
+            inputs=[feedback, log, response],
+            outputs=[log, response, code_display, feedback]
+        )
+
+# Test
+if __name__ == "__main__":
+    step_index = 0
+    app.launch(share=True)

.history/app_20250404175025.py

@@ -0,0 +1,335 @@
+import gradio as gr
+from pathlib import Path
+from reactagent.environment import Environment
+from reactagent.agents.agent_research import ResearchAgent
+from reactagent.runner import create_parser
+from reactagent import llm
+from reactagent.users.user import User
+import os
+import json
+
+
+# Global variables to store session state
+env = None
+agent = None
+state_example = False
+state_extract = False
+state_generate = False
+state_agent = False
+state_complete = False
+index_ex = "1"
+
+example_text = [
+    "Research Paper 1: Dataset and Baseline for Automatic Student Feedback Analysis",
+    "Research Paper 2: An Empirical Study on the Impact of Code Review on Software Quality"
+]
+
+
+# Load example JSON file
+def load_example_data():
+    with open("example/example_data.json", "r") as json_file:
+        example_data = json.load(json_file)
+    
+    for idx in example_data.keys():
+        try:
+            file = example_data[idx]["code_init"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_init"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")
+        try:
+            file = example_data[idx]["code_final"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_final"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")     
+    return example_data
+
+example_data = load_example_data()
+
+# Function to handle the selection of an example and populate the respective fields
+def load_example(example_id):
+    global index_ex 
+    index_ex = str(example_id)
+    example = example_data[index_ex]
+    paper_text = 'Title:\t' +  example['title'] + '\n\nAbstract:\t' + example['abstract']
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return [paper_text, tasks, gaps, keywords, recent_works]
+
+example_text = [load_example(1)[0], load_example(2)[0]]
+
+
+# Function to handle example clicks
+def load_example_and_set_index(paper_text_input):
+    global index_ex, state_example
+    state_example = True
+    index_ex = str(example_text.index(paper_text_input) + 1)
+    example =  load_example(index_ex)
+
+    return example
+
+
+
+########## Phase 1 ##############
+
+def extract_research_elements(paper_text):
+    global state_extract, index_ex, state_example
+    if not state_example or paper_text == "":
+        return "", "", "", ""
+    state_extract = True
+    if not paper_text.strip().startswith("Title:\t" + example_data[index_ex]["title"]):
+        print("Mismatch detected.")
+        print(paper_text)
+        return "", "", "", ""
+    example = example_data[index_ex]
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return tasks, gaps, keywords, recent_works
+
+
+# Step 2: Generate Research Hypothesis and Experiment Plan
+def generate_and_store(paper_text, tasks, gaps, keywords, recent_works):
+    if (not state_extract or not state_example):
+        return "", "", "", ""
+    global state_generate, index_ex
+    state_generate = True
+    hypothesis = example_data[index_ex]['hypothesis']
+    experiment_plan = example_data[index_ex]['experiment_plan']
+    return hypothesis, experiment_plan, hypothesis, experiment_plan
+
+##########  Phase 2 & 3  ##############
+def start_experiment_agent(hypothesis, plan):
+    if (not state_extract or not state_generate or not state_example):
+        return "", "", ""
+    global state_agent, step_index, state_complete
+    state_agent = True
+    step_index = 0
+    state_complete = False
+    # predefined_message = f"Implement the following hypothesis and experiment plan:\n\nHypothesis:\n{hypothesis}\n\nExperiment Plan:\n{plan}"
+    return example_data[index_ex]['code_init'], predefined_action_log, "", ""
+
+def submit_feedback(user_feedback, history, previous_response):
+    if (not state_extract or not  state_generate or not state_agent or not state_example):
+        return "", "", ""
+    global step_index, state_complete
+    step_index += 1
+    msg = history
+    if step_index < len(process_steps):
+        msg += previous_response + "\nUser feedback:" + user_feedback + "\n\n"
+        response_info = process_steps[step_index]
+        response = info_to_message(response_info)  # Convert dictionary to formatted string
+        response += "Please provide feedback based on the history, response entries, and observation, and questions: "
+        step_index += 1
+        msg += response
+    else:
+        state_complete = True
+        response = "Agent Finished."
+        
+    return msg, response, example_data[index_ex]['code_init'] if state_complete else example_data[index_ex]['code_final'], ""
+
+def load_phase_2_inputs(hypothesis, plan):
+    return hypothesis, plan, "# Code implementation will be displayed here after Start ExperimentAgent."
+
+
+
+predefined_action_log = """
+[Reasoning]: To understand the initial structure and functionality of train.py for effective improvements.
+[Action]: Inspect Script (train.py)
+Input: {"script_name": "train.py", "start_line_number": "1", "end_line_number": "74"}
+Objective: Understand the training script, including data processing, [...]
+[Observation]: The train.py script imports [...]. Sets random seeds [...]. Defines [...] Placeholder functions [...] exist without implementation. [...]
+[Feedback]: The script structure is clear, but key functions (train_model, predict) need proper implementation for proposed model training and prediction.\n
+"""
+
+
+predefined_observation = """
+Epoch [1/10],
+Train MSE: 0.543,
+Test MSE: 0.688
+Epoch [2/10],
+Train MSE: 0.242,
+Test MSE: 0.493\n
+"""
+
+# Initialize the global step_index and history
+process_steps = [
+    {
+        "Action": "Inspect Script Lines (train.py)",
+        "Observation": (
+            "The train.py script imports necessary libraries (e.g., pandas, sklearn, torch). "
+            "Sets random seeds for reproducibility. Defines compute_metrics_for_regression function "
+            "to calculate RMSE for different dimensions. Placeholder functions train_model and "
+            "predict exist without implementations."
+        ),
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The script executed successfully. Generated embeddings using the BERT model. Completed "
+            "the training process without errors. Metrics calculation placeholders indicated areas needing implementation."
+        ),
+    },
+    {
+        "Action": "Edit Script (train.py)",
+        "Observation": (
+            "Edited train.py to separate data loading, model definition, training loop, and evaluation into distinct functions. "
+            "The edited train.py now has clearly defined functions"
+            "for data loading (load_data), model definition (build_model), "
+            "training (train_model), and evaluation (evaluate_model). Similarly, eval.py is reorganized to load the model and perform predictions efficiently."
+        ),
+    },
+    {
+        "Action": "Retrieve Model",
+        "Observation": "CNN and BiLSTM retrieved.",
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The model trained over the specified number of epochs. Training and validation loss values are recorded for each epoch, "
+            "the decrease in loss indicates improved model performance."
+        )
+    },
+    {
+        "Action": "Evaluation",
+        "Observation": predefined_observation,
+    }
+]
+def info_to_message(info):
+    msg = ""
+    for k, v in info.items():
+        if isinstance(v, dict):
+            tempv = v
+            v = ""
+            for k2, v2 in tempv.items():
+                v += f"{k2}:\n  {v2}\n"
+        v = User.indent_text(v, 2)
+        msg += '-' * 64
+        msg += '\n'
+        msg += f"{k}:\n{v}\n"
+    return msg
+
+
+def handle_example_click(example_index):
+    global index_ex
+    index_ex = example_index
+    return load_example(index_ex)  # Simply return the text to display it in the textbox
+
+# Gradio Interface
+with gr.Blocks(css=".gr-examples-label {display: none;}", theme=gr.themes.Default()) as app:
+    gr.Markdown("# MLR- Copilot: Machine Learning Research based on LLM Agents")
+    gr.Markdown("###   ")
+    gr.Markdown("## <span style='color:Orange;'> This UI is for predefined example demo only.</span>")
+    gr.Markdown("## <span style='color:Orange;'> To reproduce the results please use [Github](https://github.com/du-nlp-lab/MLR-Copilot/)</span>")
+
+
+
+    gr.Markdown("MLR-Copilot is a framework where LLMs mimic researchers’ thought processes, designed to enhance the productivity of machine learning research by automating the generation and implementation of research ideas. It begins with a research paper, autonomously generating and validating these ideas, while incorporating human feedback to help reach executable research outcomes.")
+
+
+    gr.Markdown("##    ⬇️  Click an example at bottom to start！")
+
+
+    # Use state variables to store generated hypothesis and experiment plan
+    hypothesis_state = gr.State("")
+    experiment_plan_state = gr.State("")
+
+    ########## Phase 1:  Research Idea Generation Tab ##############
+    with gr.Tab("💡Stage 1: Research Idea Generation"):
+        gr.Markdown("### Extract Research Elements and Generate Research Ideas")
+        with gr.Row():
+            with gr.Column():
+                paper_text_input = gr.Textbox(value="", lines=10, label="📑 Research Paper Text", interactive=False)
+
+                # extract_button = gr.Button("🔍 Extract Research Elements")
+                with gr.Row():
+                    tasks_output = gr.Textbox(placeholder="Research task definition", label="Research Tasks", lines=2, interactive=False)
+                    gaps_output = gr.Textbox(placeholder="Research gaps of current works", label="Research Gaps", lines=2, interactive=False)
+                    keywords_output = gr.Textbox(placeholder="Paper keywords", label="Keywords", lines=2, interactive=False)
+                    recent_works_output = gr.Textbox(placeholder="Recent works extracted from Semantic Scholar", label="Recent Works", lines=2, interactive=False)
+            with gr.Column():
+                with gr.Row():  # Move the button to the top 
+                    generate_button = gr.Button("✍️ Generate Research Hypothesis & Experiment Plan")
+                with gr.Group():
+                    gr.Markdown("### 🌟 Research Idea")
+                    with gr.Row():
+                        hypothesis_output = gr.Textbox(label="Generated Hypothesis", lines=20, interactive=False)
+                        experiment_plan_output = gr.Textbox(label="Generated Experiment Plan", lines=20, interactive=False)
+        
+    
+        # Step 1: Extract Research Elements
+        # extract_button.click(
+        #     fn=extract_research_elements, 
+        #     inputs=paper_text_input, 
+        #     outputs=[tasks_output, gaps_output, keywords_output, recent_works_output]
+        # )
+
+        generate_button.click(
+            fn=generate_and_store, 
+            inputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output], 
+            outputs=[hypothesis_output, experiment_plan_output, hypothesis_state, experiment_plan_state]
+        )
+
+        gr.Examples(
+            examples=example_text,
+            inputs=[paper_text_input],
+            outputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output],
+            fn=load_example_and_set_index,
+            run_on_click = True,
+            # label="⬇️ Click an example to load"
+        )
+        
+
+
+
+    ########## Phase 2 & 3:  Experiment implementation and execution ##############
+    with gr.Tab("🧪 Stage 2 & Stage 3: Experiment implementation and execution"):
+        gr.Markdown("### Interact with the ExperimentAgent")
+
+        with gr.Row():
+            with gr.Column():
+                with gr.Group():
+                    gr.Markdown("### 🌟 Generated Research Idea")
+                    with gr.Row():
+                        idea_input = gr.Textbox(label="Generated Research Hypothesis", lines=30, interactive=False)
+                        plan_input = gr.Textbox(label="Generated Experiment Plan", lines=30, interactive=False)
+        
+            with gr.Column():
+                start_exp_agnet = gr.Button("⚙️ Start / Reset ExperimentAgent", elem_classes=["agent-btn"])
+                with gr.Group():
+                    gr.Markdown("### Implementation + Execution Log")    
+                    log = gr.Textbox(label="📖 Execution Log", lines=20, interactive=False)
+                    code_display = gr.Code(label="🧑‍💻 Implementation", language="python", interactive=False)
+
+            with gr.Column():
+                response = gr.Textbox(label="🤖 ExperimentAgent Response", lines=30, interactive=False)
+                feedback = gr.Textbox(placeholder="N/A", label="🧑‍🔬 User Feedback", lines=3, interactive=True)
+                submit_button = gr.Button("Submit", elem_classes=["Submit-btn"])
+        
+        hypothesis_state.change(
+            fn=load_phase_2_inputs, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[idea_input, plan_input, code_display]
+        )
+        
+        # Start research agent
+        start_exp_agnet.click(
+            fn=start_experiment_agent, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[code_display, log, response, feedback]
+        )
+        
+        submit_button.click(
+            fn=submit_feedback,
+            inputs=[feedback, log, response],
+            outputs=[log, response, code_display, feedback]
+        )
+
+# Test
+if __name__ == "__main__":
+    step_index = 0
+    app.launch(share=True)

.history/app_20250404175029.py

@@ -0,0 +1,335 @@
+import gradio as gr
+from pathlib import Path
+from reactagent.environment import Environment
+from reactagent.agents.agent_research import ResearchAgent
+from reactagent.runner import create_parser
+from reactagent import llm
+from reactagent.users.user import User
+import os
+import json
+
+
+# Global variables to store session state
+env = None
+agent = None
+state_example = False
+state_extract = False
+state_generate = False
+state_agent = False
+state_complete = False
+index_ex = "1"
+
+example_text = [
+    "Research Paper 1: Dataset and Baseline for Automatic Student Feedback Analysis",
+    "Research Paper 2: An Empirical Study on the Impact of Code Review on Software Quality"
+]
+
+
+# Load example JSON file
+def load_example_data():
+    with open("example/example_data.json", "r") as json_file:
+        example_data = json.load(json_file)
+    
+    for idx in example_data.keys():
+        try:
+            file = example_data[idx]["code_init"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_init"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")
+        try:
+            file = example_data[idx]["code_final"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_final"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")     
+    return example_data
+
+example_data = load_example_data()
+
+# Function to handle the selection of an example and populate the respective fields
+def load_example(example_id):
+    global index_ex 
+    index_ex = str(example_id)
+    example = example_data[index_ex]
+    paper_text = 'Title:\t' +  example['title'] + '\n\nAbstract:\t' + example['abstract']
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return [paper_text, tasks, gaps, keywords, recent_works]
+
+example_text = [load_example(1)[0], load_example(2)[0]]
+
+
+# Function to handle example clicks
+def load_example_and_set_index(paper_text_input):
+    global index_ex, state_example
+    state_example = True
+    index_ex = str(example_text.index(paper_text_input) + 1)
+    example =  load_example(index_ex)
+
+    return example
+
+
+
+########## Phase 1 ##############
+
+def extract_research_elements(paper_text):
+    global state_extract, index_ex, state_example
+    if not state_example or paper_text == "":
+        return "", "", "", ""
+    state_extract = True
+    if not paper_text.strip().startswith("Title:\t" + example_data[index_ex]["title"]):
+        print("Mismatch detected.")
+        print(paper_text)
+        return "", "", "", ""
+    example = example_data[index_ex]
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return tasks, gaps, keywords, recent_works
+
+
+# Step 2: Generate Research Hypothesis and Experiment Plan
+def generate_and_store(paper_text, tasks, gaps, keywords, recent_works):
+    if (not state_extract or not state_example):
+        return "", "", "", ""
+    global state_generate, index_ex
+    state_generate = True
+    hypothesis = example_data[index_ex]['hypothesis']
+    experiment_plan = example_data[index_ex]['experiment_plan']
+    return hypothesis, experiment_plan, hypothesis, experiment_plan
+
+##########  Phase 2 & 3  ##############
+def start_experiment_agent(hypothesis, plan):
+    if (not state_extract or not state_generate or not state_example):
+        return "", "", ""
+    global state_agent, step_index, state_complete
+    state_agent = True
+    step_index = 0
+    state_complete = False
+    # predefined_message = f"Implement the following hypothesis and experiment plan:\n\nHypothesis:\n{hypothesis}\n\nExperiment Plan:\n{plan}"
+    return example_data[index_ex]['code_init'], predefined_action_log, "", ""
+
+def submit_feedback(user_feedback, history, previous_response):
+    if (not state_extract or not  state_generate or not state_agent or not state_example):
+        return "", "", ""
+    global step_index, state_complete
+    step_index += 1
+    msg = history
+    if step_index < len(process_steps):
+        msg += previous_response + "\nUser feedback:" + user_feedback + "\n\n"
+        response_info = process_steps[step_index]
+        response = info_to_message(response_info)  # Convert dictionary to formatted string
+        response += "Please provide feedback based on the history, response entries, and observation, and questions: "
+        step_index += 1
+        msg += response
+    else:
+        state_complete = True
+        response = "Agent Finished."
+        
+    return msg, response, example_data[index_ex]['code_init'] if state_complete else example_data[index_ex]['code_final'], ""
+
+def load_phase_2_inputs(hypothesis, plan):
+    return hypothesis, plan, "# Code implementation will be displayed here after Start ExperimentAgent."
+
+
+
+predefined_action_log = """
+[Reasoning]: To understand the initial structure and functionality of train.py for effective improvements.
+[Action]: Inspect Script (train.py)
+Input: {"script_name": "train.py", "start_line_number": "1", "end_line_number": "74"}
+Objective: Understand the training script, including data processing, [...]
+[Observation]: The train.py script imports [...]. Sets random seeds [...]. Defines [...] Placeholder functions [...] exist without implementation. [...]
+[Feedback]: The script structure is clear, but key functions (train_model, predict) need proper implementation for proposed model training and prediction.\n
+"""
+
+
+predefined_observation = """
+Epoch [1/10],
+Train MSE: 0.543,
+Test MSE: 0.688
+Epoch [2/10],
+Train MSE: 0.242,
+Test MSE: 0.493\n
+"""
+
+# Initialize the global step_index and history
+process_steps = [
+    {
+        "Action": "Inspect Script Lines (train.py)",
+        "Observation": (
+            "The train.py script imports necessary libraries (e.g., pandas, sklearn, torch). "
+            "Sets random seeds for reproducibility. Defines compute_metrics_for_regression function "
+            "to calculate RMSE for different dimensions. Placeholder functions train_model and "
+            "predict exist without implementations."
+        ),
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The script executed successfully. Generated embeddings using the BERT model. Completed "
+            "the training process without errors. Metrics calculation placeholders indicated areas needing implementation."
+        ),
+    },
+    {
+        "Action": "Edit Script (train.py)",
+        "Observation": (
+            "Edited train.py to separate data loading, model definition, training loop, and evaluation into distinct functions. "
+            "The edited train.py now has clearly defined functions"
+            "for data loading (load_data), model definition (build_model), "
+            "training (train_model), and evaluation (evaluate_model). Similarly, eval.py is reorganized to load the model and perform predictions efficiently."
+        ),
+    },
+    {
+        "Action": "Retrieve Model",
+        "Observation": "CNN and BiLSTM retrieved.",
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The model trained over the specified number of epochs. Training and validation loss values are recorded for each epoch, "
+            "the decrease in loss indicates improved model performance."
+        )
+    },
+    {
+        "Action": "Evaluation",
+        "Observation": predefined_observation,
+    }
+]
+def info_to_message(info):
+    msg = ""
+    for k, v in info.items():
+        if isinstance(v, dict):
+            tempv = v
+            v = ""
+            for k2, v2 in tempv.items():
+                v += f"{k2}:\n  {v2}\n"
+        v = User.indent_text(v, 2)
+        msg += '-' * 64
+        msg += '\n'
+        msg += f"{k}:\n{v}\n"
+    return msg
+
+
+def handle_example_click(example_index):
+    global index_ex
+    index_ex = example_index
+    return load_example(index_ex)  # Simply return the text to display it in the textbox
+
+# Gradio Interface
+with gr.Blocks(css=".gr-examples-label {display: none;}", theme=gr.themes.Default()) as app:
+    gr.Markdown("# MLR- Copilot: Machine Learning Research based on LLM Agents")
+    gr.Markdown("###   ")
+    gr.Markdown("## <span style='color:Orange;'> This UI is for predefined example demo only.</span>")
+    gr.Markdown("## <span style='color:Orange;'> To reproduce the results please use [Github](https://github.com/du-nlp-lab/MLR-Copilot/)</span>")
+
+
+
+    gr.Markdown("MLR-Copilot is a framework where LLMs mimic researchers’ thought processes, designed to enhance the productivity of machine learning research by automating the generation and implementation of research ideas. It begins with a research paper, autonomously generating and validating these ideas, while incorporating human feedback to help reach executable research outcomes.")
+
+
+    gr.Markdown("##    ⬇️  Click an example at bottom to start！")
+
+
+    # Use state variables to store generated hypothesis and experiment plan
+    hypothesis_state = gr.State("")
+    experiment_plan_state = gr.State("")
+
+    ########## Phase 1:  Research Idea Generation Tab ##############
+    with gr.Tab("💡Stage 1: Research Idea Generation"):
+        gr.Markdown("### Extract Research Elements and Generate Research Ideas")
+        with gr.Row():
+            with gr.Column():
+                paper_text_input = gr.Textbox(value="", lines=10, label="📑 Research Paper Text", interactive=False)
+
+                # extract_button = gr.Button("🔍 Extract Research Elements")
+                with gr.Row():
+                    tasks_output = gr.Textbox(placeholder="Research task definition", label="Research Tasks", lines=2, interactive=False)
+                    gaps_output = gr.Textbox(placeholder="Research gaps of current works", label="Research Gaps", lines=2, interactive=False)
+                    keywords_output = gr.Textbox(placeholder="Paper keywords", label="Keywords", lines=2, interactive=False)
+                    recent_works_output = gr.Textbox(placeholder="Recent works extracted from Semantic Scholar", label="Recent Works", lines=2, interactive=False)
+            with gr.Column():
+                with gr.Row():  # Move the button to the top 
+                    generate_button = gr.Button("✍️ Generate Research Hypothesis & Experiment Plan")
+                with gr.Group():
+                    gr.Markdown("### 🌟 Research Idea")
+                    with gr.Row():
+                        hypothesis_output = gr.Textbox(label="Generated Hypothesis", lines=20, interactive=False)
+                        experiment_plan_output = gr.Textbox(label="Generated Experiment Plan", lines=20, interactive=False)
+        
+    
+        # Step 1: Extract Research Elements
+        # extract_button.click(
+        #     fn=extract_research_elements, 
+        #     inputs=paper_text_input, 
+        #     outputs=[tasks_output, gaps_output, keywords_output, recent_works_output]
+        # )
+
+        generate_button.click(
+            fn=generate_and_store, 
+            inputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output], 
+            outputs=[hypothesis_output, experiment_plan_output, hypothesis_state, experiment_plan_state]
+        )
+
+        gr.Examples(
+            examples=example_text,
+            inputs=[paper_text_input],
+            outputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output],
+            fn=load_example_and_set_index,
+            run_on_click = True,
+            # label="⬇️ Click an example to load"
+        )
+        
+
+
+
+    ########## Phase 2 & 3:  Experiment implementation and execution ##############
+    with gr.Tab("🧪 Stage 2 & Stage 3: Experiment implementation and execution"):
+        gr.Markdown("### Interact with the ExperimentAgent")
+
+        with gr.Row():
+            with gr.Column():
+                with gr.Group():
+                    gr.Markdown("### 🌟 Generated Research Idea")
+                    with gr.Row():
+                        idea_input = gr.Textbox(label="Generated Research Hypothesis", lines=30, interactive=False)
+                        plan_input = gr.Textbox(label="Generated Experiment Plan", lines=30, interactive=False)
+        
+            with gr.Column():
+                start_exp_agnet = gr.Button("⚙️ Start / Reset ExperimentAgent", elem_classes=["agent-btn"])
+                with gr.Group():
+                    gr.Markdown("### Implementation + Execution Log")    
+                    log = gr.Textbox(label="📖 Execution Log", lines=20, interactive=False)
+                    code_display = gr.Code(label="🧑‍💻 Implementation", language="python", interactive=False)
+
+            with gr.Column():
+                response = gr.Textbox(label="🤖 ExperimentAgent Response", lines=30, interactive=False)
+                feedback = gr.Textbox(placeholder="N/A", label="🧑‍🔬 User Feedback", lines=3, interactive=True)
+                submit_button = gr.Button("Submit", elem_classes=["Submit-btn"])
+        
+        hypothesis_state.change(
+            fn=load_phase_2_inputs, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[idea_input, plan_input, code_display]
+        )
+        
+        # Start research agent
+        start_exp_agnet.click(
+            fn=start_experiment_agent, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[code_display, log, response, feedback]
+        )
+        
+        submit_button.click(
+            fn=submit_feedback,
+            inputs=[feedback, log, response],
+            outputs=[log, response, code_display, feedback]
+        )
+
+# Test
+if __name__ == "__main__":
+    step_index = 0
+    app.launch(share=True)

.history/app_20250404175030.py

@@ -0,0 +1,335 @@
+import gradio as gr
+from pathlib import Path
+from reactagent.environment import Environment
+from reactagent.agents.agent_research import ResearchAgent
+from reactagent.runner import create_parser
+from reactagent import llm
+from reactagent.users.user import User
+import os
+import json
+
+
+# Global variables to store session state
+env = None
+agent = None
+state_example = False
+state_extract = False
+state_generate = False
+state_agent = False
+state_complete = False
+index_ex = "1"
+
+example_text = [
+    "Research Paper 1: Dataset and Baseline for Automatic Student Feedback Analysis",
+    "Research Paper 2: An Empirical Study on the Impact of Code Review on Software Quality"
+]
+
+
+# Load example JSON file
+def load_example_data():
+    with open("example/example_data.json", "r") as json_file:
+        example_data = json.load(json_file)
+    
+    for idx in example_data.keys():
+        try:
+            file = example_data[idx]["code_init"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_init"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")
+        try:
+            file = example_data[idx]["code_final"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_final"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")     
+    return example_data
+
+example_data = load_example_data()
+
+# Function to handle the selection of an example and populate the respective fields
+def load_example(example_id):
+    global index_ex 
+    index_ex = str(example_id)
+    example = example_data[index_ex]
+    paper_text = 'Title:\t' +  example['title'] + '\n\nAbstract:\t' + example['abstract']
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return [paper_text, tasks, gaps, keywords, recent_works]
+
+example_text = [load_example(1)[0], load_example(2)[0]]
+
+
+# Function to handle example clicks
+def load_example_and_set_index(paper_text_input):
+    global index_ex, state_example
+    state_example = True
+    index_ex = str(example_text.index(paper_text_input) + 1)
+    example =  load_example(index_ex)
+
+    return example
+
+
+
+########## Phase 1 ##############
+
+def extract_research_elements(paper_text):
+    global state_extract, index_ex, state_example
+    if not state_example or paper_text == "":
+        return "", "", "", ""
+    state_extract = True
+    if not paper_text.strip().startswith("Title:\t" + example_data[index_ex]["title"]):
+        print("Mismatch detected.")
+        print(paper_text)
+        return "", "", "", ""
+    example = example_data[index_ex]
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return tasks, gaps, keywords, recent_works
+
+
+# Step 2: Generate Research Hypothesis and Experiment Plan
+def generate_and_store(paper_text, tasks, gaps, keywords, recent_works):
+    if (not state_extract or not state_example):
+        return "", "", "", ""
+    global state_generate, index_ex
+    state_generate = True
+    hypothesis = example_data[index_ex]['hypothesis']
+    experiment_plan = example_data[index_ex]['experiment_plan']
+    return hypothesis, experiment_plan, hypothesis, experiment_plan
+
+##########  Phase 2 & 3  ##############
+def start_experiment_agent(hypothesis, plan):
+    if (not state_extract or not state_generate or not state_example):
+        return "", "", ""
+    global state_agent, step_index, state_complete
+    state_agent = True
+    step_index = 0
+    state_complete = False
+    # predefined_message = f"Implement the following hypothesis and experiment plan:\n\nHypothesis:\n{hypothesis}\n\nExperiment Plan:\n{plan}"
+    return example_data[index_ex]['code_init'], predefined_action_log, "", ""
+
+def submit_feedback(user_feedback, history, previous_response):
+    if (not state_extract or not  state_generate or not state_agent or not state_example):
+        return "", "", ""
+    global step_index, state_complete
+    step_index += 1
+    msg = history
+    if step_index < len(process_steps):
+        msg += previous_response + "\nUser feedback:" + user_feedback + "\n\n"
+        response_info = process_steps[step_index]
+        response = info_to_message(response_info)  # Convert dictionary to formatted string
+        response += "Please provide feedback based on the history, response entries, and observation, and questions: "
+        step_index += 1
+        msg += response
+    else:
+        state_complete = True
+        response = "Agent Finished."
+        
+    return msg, response, example_data[index_ex]['code_init'] if state_complete else example_data[index_ex]['code_final'], ""
+
+def load_phase_2_inputs(hypothesis, plan):
+    return hypothesis, plan, "# Code implementation will be displayed here after Start ExperimentAgent."
+
+
+
+predefined_action_log = """
+[Reasoning]: To understand the initial structure and functionality of train.py for effective improvements.
+[Action]: Inspect Script (train.py)
+Input: {"script_name": "train.py", "start_line_number": "1", "end_line_number": "74"}
+Objective: Understand the training script, including data processing, [...]
+[Observation]: The train.py script imports [...]. Sets random seeds [...]. Defines [...] Placeholder functions [...] exist without implementation. [...]
+[Feedback]: The script structure is clear, but key functions (train_model, predict) need proper implementation for proposed model training and prediction.\n
+"""
+
+
+predefined_observation = """
+Epoch [1/10],
+Train MSE: 0.543,
+Test MSE: 0.688
+Epoch [2/10],
+Train MSE: 0.242,
+Test MSE: 0.493\n
+"""
+
+# Initialize the global step_index and history
+process_steps = [
+    {
+        "Action": "Inspect Script Lines (train.py)",
+        "Observation": (
+            "The train.py script imports necessary libraries (e.g., pandas, sklearn, torch). "
+            "Sets random seeds for reproducibility. Defines compute_metrics_for_regression function "
+            "to calculate RMSE for different dimensions. Placeholder functions train_model and "
+            "predict exist without implementations."
+        ),
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The script executed successfully. Generated embeddings using the BERT model. Completed "
+            "the training process without errors. Metrics calculation placeholders indicated areas needing implementation."
+        ),
+    },
+    {
+        "Action": "Edit Script (train.py)",
+        "Observation": (
+            "Edited train.py to separate data loading, model definition, training loop, and evaluation into distinct functions. "
+            "The edited train.py now has clearly defined functions"
+            "for data loading (load_data), model definition (build_model), "
+            "training (train_model), and evaluation (evaluate_model). Similarly, eval.py is reorganized to load the model and perform predictions efficiently."
+        ),
+    },
+    {
+        "Action": "Retrieve Model",
+        "Observation": "CNN and BiLSTM retrieved.",
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The model trained over the specified number of epochs. Training and validation loss values are recorded for each epoch, "
+            "the decrease in loss indicates improved model performance."
+        )
+    },
+    {
+        "Action": "Evaluation",
+        "Observation": predefined_observation,
+    }
+]
+def info_to_message(info):
+    msg = ""
+    for k, v in info.items():
+        if isinstance(v, dict):
+            tempv = v
+            v = ""
+            for k2, v2 in tempv.items():
+                v += f"{k2}:\n  {v2}\n"
+        v = User.indent_text(v, 2)
+        msg += '-' * 64
+        msg += '\n'
+        msg += f"{k}:\n{v}\n"
+    return msg
+
+
+def handle_example_click(example_index):
+    global index_ex
+    index_ex = example_index
+    return load_example(index_ex)  # Simply return the text to display it in the textbox
+
+# Gradio Interface
+with gr.Blocks(css=".gr-examples-label {display: none;}", theme=gr.themes.Default()) as app:
+    gr.Markdown("# MLR- Copilot: Machine Learning Research based on LLM Agents")
+    gr.Markdown("###   ")
+    gr.Markdown("## <span style='color:Orange;'> This UI is for predefined example demo only.</span>")
+    gr.Markdown("## <span style='color:Orange;'> To reproduce the results please use [Github](https://github.com/du-nlp-lab/MLR-Copilot/)</span>")
+
+
+
+    gr.Markdown("MLR-Copilot is a framework where LLMs mimic researchers’ thought processes, designed to enhance the productivity of machine learning research by automating the generation and implementation of research ideas. It begins with a research paper, autonomously generating and validating these ideas, while incorporating human feedback to help reach executable research outcomes.")
+
+
+    gr.Markdown("##    ⬇️  Click an example at bottom to start！")
+
+
+    # Use state variables to store generated hypothesis and experiment plan
+    hypothesis_state = gr.State("")
+    experiment_plan_state = gr.State("")
+
+    ########## Phase 1:  Research Idea Generation Tab ##############
+    with gr.Tab("💡Stage 1: Research Idea Generation"):
+        gr.Markdown("### Extract Research Elements and Generate Research Ideas")
+        with gr.Row():
+            with gr.Column():
+                paper_text_input = gr.Textbox(value="", lines=10, label="📑 Research Paper Text", interactive=False)
+
+                # extract_button = gr.Button("🔍 Extract Research Elements")
+                with gr.Row():
+                    tasks_output = gr.Textbox(placeholder="Research task definition", label="Research Tasks", lines=2, interactive=False)
+                    gaps_output = gr.Textbox(placeholder="Research gaps of current works", label="Research Gaps", lines=2, interactive=False)
+                    keywords_output = gr.Textbox(placeholder="Paper keywords", label="Keywords", lines=2, interactive=False)
+                    recent_works_output = gr.Textbox(placeholder="Recent works extracted from Semantic Scholar", label="Recent Works", lines=2, interactive=False)
+            with gr.Column():
+                with gr.Row():  # Move the button to the top 
+                    generate_button = gr.Button("✍️ Generate Research Hypothesis & Experiment Plan")
+                with gr.Group():
+                    gr.Markdown("### 🌟 Research Idea")
+                    with gr.Row():
+                        hypothesis_output = gr.Textbox(label="Generated Hypothesis", lines=20, interactive=False)
+                        experiment_plan_output = gr.Textbox(label="Generated Experiment Plan", lines=20, interactive=False)
+        
+    
+        # Step 1: Extract Research Elements
+        # extract_button.click(
+        #     fn=extract_research_elements, 
+        #     inputs=paper_text_input, 
+        #     outputs=[tasks_output, gaps_output, keywords_output, recent_works_output]
+        # )
+
+        generate_button.click(
+            fn=generate_and_store, 
+            inputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output], 
+            outputs=[hypothesis_output, experiment_plan_output, hypothesis_state, experiment_plan_state]
+        )
+
+        gr.Examples(
+            examples=example_text,
+            inputs=[paper_text_input],
+            outputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output],
+            fn=load_example_and_set_index,
+            run_on_click = True,
+            # label="⬇️ Click an example to load"
+        )
+        
+
+
+
+    ########## Phase 2 & 3:  Experiment implementation and execution ##############
+    with gr.Tab("🧪 Stage 2 & Stage 3: Experiment implementation and execution"):
+        gr.Markdown("### Interact with the ExperimentAgent")
+
+        with gr.Row():
+            with gr.Column():
+                with gr.Group():
+                    gr.Markdown("### 🌟 Generated Research Idea")
+                    with gr.Row():
+                        idea_input = gr.Textbox(label="Generated Research Hypothesis", lines=30, interactive=False)
+                        plan_input = gr.Textbox(label="Generated Experiment Plan", lines=30, interactive=False)
+        
+            with gr.Column():
+                start_exp_agnet = gr.Button("⚙️ Start / Reset ExperimentAgent", elem_classes=["agent-btn"])
+                with gr.Group():
+                    gr.Markdown("### Implementation + Execution Log")    
+                    log = gr.Textbox(label="📖 Execution Log", lines=20, interactive=False)
+                    code_display = gr.Code(label="🧑‍💻 Implementation", language="python", interactive=False)
+
+            with gr.Column():
+                response = gr.Textbox(label="🤖 ExperimentAgent Response", lines=30, interactive=False)
+                feedback = gr.Textbox(placeholder="N/A", label="🧑‍🔬 User Feedback", lines=3, interactive=True)
+                submit_button = gr.Button("Submit", elem_classes=["Submit-btn"])
+        
+        hypothesis_state.change(
+            fn=load_phase_2_inputs, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[idea_input, plan_input, code_display]
+        )
+        
+        # Start research agent
+        start_exp_agnet.click(
+            fn=start_experiment_agent, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[code_display, log, response, feedback]
+        )
+        
+        submit_button.click(
+            fn=submit_feedback,
+            inputs=[feedback, log, response],
+            outputs=[log, response, code_display, feedback]
+        )
+
+# Test
+if __name__ == "__main__":
+    step_index = 0
+    app.launch(share=True)

.history/app_20250404175127.py

@@ -0,0 +1,335 @@
+import gradio as gr
+from pathlib import Path
+from reactagent.environment import Environment
+from reactagent.agents.agent_research import ResearchAgent
+from reactagent.runner import create_parser
+from reactagent import llm
+from reactagent.users.user import User
+import os
+import json
+
+
+# Global variables to store session state
+env = None
+agent = None
+state_example = False
+state_extract = False
+state_generate = False
+state_agent = False
+state_complete = False
+index_ex = "1"
+
+example_text = [
+    "Research Paper 1: Dataset and Baseline for Automatic Student Feedback Analysis",
+    "Research Paper 2: An Empirical Study on the Impact of Code Review on Software Quality"
+]
+
+
+# Load example JSON file
+def load_example_data():
+    with open("example/example_data.json", "r") as json_file:
+        example_data = json.load(json_file)
+    
+    for idx in example_data.keys():
+        try:
+            file = example_data[idx]["code_init"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_init"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")
+        try:
+            file = example_data[idx]["code_final"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_final"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")     
+    return example_data
+
+example_data = load_example_data()
+
+# Function to handle the selection of an example and populate the respective fields
+def load_example(example_id):
+    global index_ex 
+    index_ex = str(example_id)
+    example = example_data[index_ex]
+    paper_text = 'Title:\t' +  example['title'] + '\n\nAbstract:\t' + example['abstract']
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return [paper_text, tasks, gaps, keywords, recent_works]
+
+example_text = [load_example(1)[0], load_example(2)[0]]
+
+
+# Function to handle example clicks
+def load_example_and_set_index(paper_text_input):
+    global index_ex, state_example
+    state_example = True
+    index_ex = str(example_text.index(paper_text_input) + 1)
+    example =  load_example(index_ex)
+
+    return example
+
+
+
+########## Phase 1 ##############
+
+def extract_research_elements(paper_text):
+    global state_extract, index_ex, state_example
+    if not state_example or paper_text == "":
+        return "", "", "", ""
+    state_extract = True
+    if not paper_text.strip().startswith("Title:\t" + example_data[index_ex]["title"]):
+        print("Mismatch detected.")
+        print(paper_text)
+        return "", "", "", ""
+    example = example_data[index_ex]
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return tasks, gaps, keywords, recent_works
+
+
+# Step 2: Generate Research Hypothesis and Experiment Plan
+def generate_and_store(paper_text, tasks, gaps, keywords, recent_works):
+    if (not state_example):
+        return "", "", "", ""
+    global state_generate, index_ex
+    state_generate = True
+    hypothesis = example_data[index_ex]['hypothesis']
+    experiment_plan = example_data[index_ex]['experiment_plan']
+    return hypothesis, experiment_plan, hypothesis, experiment_plan
+
+##########  Phase 2 & 3  ##############
+def start_experiment_agent(hypothesis, plan):
+    if (not state_extract or not state_generate or not state_example):
+        return "", "", ""
+    global state_agent, step_index, state_complete
+    state_agent = True
+    step_index = 0
+    state_complete = False
+    # predefined_message = f"Implement the following hypothesis and experiment plan:\n\nHypothesis:\n{hypothesis}\n\nExperiment Plan:\n{plan}"
+    return example_data[index_ex]['code_init'], predefined_action_log, "", ""
+
+def submit_feedback(user_feedback, history, previous_response):
+    if (not state_extract or not  state_generate or not state_agent or not state_example):
+        return "", "", ""
+    global step_index, state_complete
+    step_index += 1
+    msg = history
+    if step_index < len(process_steps):
+        msg += previous_response + "\nUser feedback:" + user_feedback + "\n\n"
+        response_info = process_steps[step_index]
+        response = info_to_message(response_info)  # Convert dictionary to formatted string
+        response += "Please provide feedback based on the history, response entries, and observation, and questions: "
+        step_index += 1
+        msg += response
+    else:
+        state_complete = True
+        response = "Agent Finished."
+        
+    return msg, response, example_data[index_ex]['code_init'] if state_complete else example_data[index_ex]['code_final'], ""
+
+def load_phase_2_inputs(hypothesis, plan):
+    return hypothesis, plan, "# Code implementation will be displayed here after Start ExperimentAgent."
+
+
+
+predefined_action_log = """
+[Reasoning]: To understand the initial structure and functionality of train.py for effective improvements.
+[Action]: Inspect Script (train.py)
+Input: {"script_name": "train.py", "start_line_number": "1", "end_line_number": "74"}
+Objective: Understand the training script, including data processing, [...]
+[Observation]: The train.py script imports [...]. Sets random seeds [...]. Defines [...] Placeholder functions [...] exist without implementation. [...]
+[Feedback]: The script structure is clear, but key functions (train_model, predict) need proper implementation for proposed model training and prediction.\n
+"""
+
+
+predefined_observation = """
+Epoch [1/10],
+Train MSE: 0.543,
+Test MSE: 0.688
+Epoch [2/10],
+Train MSE: 0.242,
+Test MSE: 0.493\n
+"""
+
+# Initialize the global step_index and history
+process_steps = [
+    {
+        "Action": "Inspect Script Lines (train.py)",
+        "Observation": (
+            "The train.py script imports necessary libraries (e.g., pandas, sklearn, torch). "
+            "Sets random seeds for reproducibility. Defines compute_metrics_for_regression function "
+            "to calculate RMSE for different dimensions. Placeholder functions train_model and "
+            "predict exist without implementations."
+        ),
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The script executed successfully. Generated embeddings using the BERT model. Completed "
+            "the training process without errors. Metrics calculation placeholders indicated areas needing implementation."
+        ),
+    },
+    {
+        "Action": "Edit Script (train.py)",
+        "Observation": (
+            "Edited train.py to separate data loading, model definition, training loop, and evaluation into distinct functions. "
+            "The edited train.py now has clearly defined functions"
+            "for data loading (load_data), model definition (build_model), "
+            "training (train_model), and evaluation (evaluate_model). Similarly, eval.py is reorganized to load the model and perform predictions efficiently."
+        ),
+    },
+    {
+        "Action": "Retrieve Model",
+        "Observation": "CNN and BiLSTM retrieved.",
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The model trained over the specified number of epochs. Training and validation loss values are recorded for each epoch, "
+            "the decrease in loss indicates improved model performance."
+        )
+    },
+    {
+        "Action": "Evaluation",
+        "Observation": predefined_observation,
+    }
+]
+def info_to_message(info):
+    msg = ""
+    for k, v in info.items():
+        if isinstance(v, dict):
+            tempv = v
+            v = ""
+            for k2, v2 in tempv.items():
+                v += f"{k2}:\n  {v2}\n"
+        v = User.indent_text(v, 2)
+        msg += '-' * 64
+        msg += '\n'
+        msg += f"{k}:\n{v}\n"
+    return msg
+
+
+def handle_example_click(example_index):
+    global index_ex
+    index_ex = example_index
+    return load_example(index_ex)  # Simply return the text to display it in the textbox
+
+# Gradio Interface
+with gr.Blocks(css=".gr-examples-label {display: none;}", theme=gr.themes.Default()) as app:
+    gr.Markdown("# MLR- Copilot: Machine Learning Research based on LLM Agents")
+    gr.Markdown("###   ")
+    gr.Markdown("## <span style='color:Orange;'> This UI is for predefined example demo only.</span>")
+    gr.Markdown("## <span style='color:Orange;'> To reproduce the results please use [Github](https://github.com/du-nlp-lab/MLR-Copilot/)</span>")
+
+
+
+    gr.Markdown("MLR-Copilot is a framework where LLMs mimic researchers’ thought processes, designed to enhance the productivity of machine learning research by automating the generation and implementation of research ideas. It begins with a research paper, autonomously generating and validating these ideas, while incorporating human feedback to help reach executable research outcomes.")
+
+
+    gr.Markdown("##    ⬇️  Click an example at bottom to start！")
+
+
+    # Use state variables to store generated hypothesis and experiment plan
+    hypothesis_state = gr.State("")
+    experiment_plan_state = gr.State("")
+
+    ########## Phase 1:  Research Idea Generation Tab ##############
+    with gr.Tab("💡Stage 1: Research Idea Generation"):
+        gr.Markdown("### Extract Research Elements and Generate Research Ideas")
+        with gr.Row():
+            with gr.Column():
+                paper_text_input = gr.Textbox(value="", lines=10, label="📑 Research Paper Text", interactive=False)
+
+                # extract_button = gr.Button("🔍 Extract Research Elements")
+                with gr.Row():
+                    tasks_output = gr.Textbox(placeholder="Research task definition", label="Research Tasks", lines=2, interactive=False)
+                    gaps_output = gr.Textbox(placeholder="Research gaps of current works", label="Research Gaps", lines=2, interactive=False)
+                    keywords_output = gr.Textbox(placeholder="Paper keywords", label="Keywords", lines=2, interactive=False)
+                    recent_works_output = gr.Textbox(placeholder="Recent works extracted from Semantic Scholar", label="Recent Works", lines=2, interactive=False)
+            with gr.Column():
+                with gr.Row():  # Move the button to the top 
+                    generate_button = gr.Button("✍️ Generate Research Hypothesis & Experiment Plan")
+                with gr.Group():
+                    gr.Markdown("### 🌟 Research Idea")
+                    with gr.Row():
+                        hypothesis_output = gr.Textbox(label="Generated Hypothesis", lines=20, interactive=False)
+                        experiment_plan_output = gr.Textbox(label="Generated Experiment Plan", lines=20, interactive=False)
+        
+    
+        # Step 1: Extract Research Elements
+        # extract_button.click(
+        #     fn=extract_research_elements, 
+        #     inputs=paper_text_input, 
+        #     outputs=[tasks_output, gaps_output, keywords_output, recent_works_output]
+        # )
+
+        generate_button.click(
+            fn=generate_and_store, 
+            inputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output], 
+            outputs=[hypothesis_output, experiment_plan_output, hypothesis_state, experiment_plan_state]
+        )
+
+        gr.Examples(
+            examples=example_text,
+            inputs=[paper_text_input],
+            outputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output],
+            fn=load_example_and_set_index,
+            run_on_click = True,
+            # label="⬇️ Click an example to load"
+        )
+        
+
+
+
+    ########## Phase 2 & 3:  Experiment implementation and execution ##############
+    with gr.Tab("🧪 Stage 2 & Stage 3: Experiment implementation and execution"):
+        gr.Markdown("### Interact with the ExperimentAgent")
+
+        with gr.Row():
+            with gr.Column():
+                with gr.Group():
+                    gr.Markdown("### 🌟 Generated Research Idea")
+                    with gr.Row():
+                        idea_input = gr.Textbox(label="Generated Research Hypothesis", lines=30, interactive=False)
+                        plan_input = gr.Textbox(label="Generated Experiment Plan", lines=30, interactive=False)
+        
+            with gr.Column():
+                start_exp_agnet = gr.Button("⚙️ Start / Reset ExperimentAgent", elem_classes=["agent-btn"])
+                with gr.Group():
+                    gr.Markdown("### Implementation + Execution Log")    
+                    log = gr.Textbox(label="📖 Execution Log", lines=20, interactive=False)
+                    code_display = gr.Code(label="🧑‍💻 Implementation", language="python", interactive=False)
+
+            with gr.Column():
+                response = gr.Textbox(label="🤖 ExperimentAgent Response", lines=30, interactive=False)
+                feedback = gr.Textbox(placeholder="N/A", label="🧑‍🔬 User Feedback", lines=3, interactive=True)
+                submit_button = gr.Button("Submit", elem_classes=["Submit-btn"])
+        
+        hypothesis_state.change(
+            fn=load_phase_2_inputs, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[idea_input, plan_input, code_display]
+        )
+        
+        # Start research agent
+        start_exp_agnet.click(
+            fn=start_experiment_agent, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[code_display, log, response, feedback]
+        )
+        
+        submit_button.click(
+            fn=submit_feedback,
+            inputs=[feedback, log, response],
+            outputs=[log, response, code_display, feedback]
+        )
+
+# Test
+if __name__ == "__main__":
+    step_index = 0
+    app.launch(share=True)

.history/app_20250404175128.py

@@ -0,0 +1,335 @@
+import gradio as gr
+from pathlib import Path
+from reactagent.environment import Environment
+from reactagent.agents.agent_research import ResearchAgent
+from reactagent.runner import create_parser
+from reactagent import llm
+from reactagent.users.user import User
+import os
+import json
+
+
+# Global variables to store session state
+env = None
+agent = None
+state_example = False
+state_extract = False
+state_generate = False
+state_agent = False
+state_complete = False
+index_ex = "1"
+
+example_text = [
+    "Research Paper 1: Dataset and Baseline for Automatic Student Feedback Analysis",
+    "Research Paper 2: An Empirical Study on the Impact of Code Review on Software Quality"
+]
+
+
+# Load example JSON file
+def load_example_data():
+    with open("example/example_data.json", "r") as json_file:
+        example_data = json.load(json_file)
+    
+    for idx in example_data.keys():
+        try:
+            file = example_data[idx]["code_init"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_init"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")
+        try:
+            file = example_data[idx]["code_final"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_final"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")     
+    return example_data
+
+example_data = load_example_data()
+
+# Function to handle the selection of an example and populate the respective fields
+def load_example(example_id):
+    global index_ex 
+    index_ex = str(example_id)
+    example = example_data[index_ex]
+    paper_text = 'Title:\t' +  example['title'] + '\n\nAbstract:\t' + example['abstract']
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return [paper_text, tasks, gaps, keywords, recent_works]
+
+example_text = [load_example(1)[0], load_example(2)[0]]
+
+
+# Function to handle example clicks
+def load_example_and_set_index(paper_text_input):
+    global index_ex, state_example
+    state_example = True
+    index_ex = str(example_text.index(paper_text_input) + 1)
+    example =  load_example(index_ex)
+
+    return example
+
+
+
+########## Phase 1 ##############
+
+def extract_research_elements(paper_text):
+    global state_extract, index_ex, state_example
+    if not state_example or paper_text == "":
+        return "", "", "", ""
+    state_extract = True
+    if not paper_text.strip().startswith("Title:\t" + example_data[index_ex]["title"]):
+        print("Mismatch detected.")
+        print(paper_text)
+        return "", "", "", ""
+    example = example_data[index_ex]
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return tasks, gaps, keywords, recent_works
+
+
+# Step 2: Generate Research Hypothesis and Experiment Plan
+def generate_and_store(paper_text, tasks, gaps, keywords, recent_works):
+    if (not state_example):
+        return "", "", "", ""
+    global state_generate, index_ex
+    state_generate = True
+    hypothesis = example_data[index_ex]['hypothesis']
+    experiment_plan = example_data[index_ex]['experiment_plan']
+    return hypothesis, experiment_plan, hypothesis, experiment_plan
+
+##########  Phase 2 & 3  ##############
+def start_experiment_agent(hypothesis, plan):
+    if (not state_extract or not state_generate or not state_example):
+        return "", "", ""
+    global state_agent, step_index, state_complete
+    state_agent = True
+    step_index = 0
+    state_complete = False
+    # predefined_message = f"Implement the following hypothesis and experiment plan:\n\nHypothesis:\n{hypothesis}\n\nExperiment Plan:\n{plan}"
+    return example_data[index_ex]['code_init'], predefined_action_log, "", ""
+
+def submit_feedback(user_feedback, history, previous_response):
+    if (not state_extract or not  state_generate or not state_agent or not state_example):
+        return "", "", ""
+    global step_index, state_complete
+    step_index += 1
+    msg = history
+    if step_index < len(process_steps):
+        msg += previous_response + "\nUser feedback:" + user_feedback + "\n\n"
+        response_info = process_steps[step_index]
+        response = info_to_message(response_info)  # Convert dictionary to formatted string
+        response += "Please provide feedback based on the history, response entries, and observation, and questions: "
+        step_index += 1
+        msg += response
+    else:
+        state_complete = True
+        response = "Agent Finished."
+        
+    return msg, response, example_data[index_ex]['code_init'] if state_complete else example_data[index_ex]['code_final'], ""
+
+def load_phase_2_inputs(hypothesis, plan):
+    return hypothesis, plan, "# Code implementation will be displayed here after Start ExperimentAgent."
+
+
+
+predefined_action_log = """
+[Reasoning]: To understand the initial structure and functionality of train.py for effective improvements.
+[Action]: Inspect Script (train.py)
+Input: {"script_name": "train.py", "start_line_number": "1", "end_line_number": "74"}
+Objective: Understand the training script, including data processing, [...]
+[Observation]: The train.py script imports [...]. Sets random seeds [...]. Defines [...] Placeholder functions [...] exist without implementation. [...]
+[Feedback]: The script structure is clear, but key functions (train_model, predict) need proper implementation for proposed model training and prediction.\n
+"""
+
+
+predefined_observation = """
+Epoch [1/10],
+Train MSE: 0.543,
+Test MSE: 0.688
+Epoch [2/10],
+Train MSE: 0.242,
+Test MSE: 0.493\n
+"""
+
+# Initialize the global step_index and history
+process_steps = [
+    {
+        "Action": "Inspect Script Lines (train.py)",
+        "Observation": (
+            "The train.py script imports necessary libraries (e.g., pandas, sklearn, torch). "
+            "Sets random seeds for reproducibility. Defines compute_metrics_for_regression function "
+            "to calculate RMSE for different dimensions. Placeholder functions train_model and "
+            "predict exist without implementations."
+        ),
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The script executed successfully. Generated embeddings using the BERT model. Completed "
+            "the training process without errors. Metrics calculation placeholders indicated areas needing implementation."
+        ),
+    },
+    {
+        "Action": "Edit Script (train.py)",
+        "Observation": (
+            "Edited train.py to separate data loading, model definition, training loop, and evaluation into distinct functions. "
+            "The edited train.py now has clearly defined functions"
+            "for data loading (load_data), model definition (build_model), "
+            "training (train_model), and evaluation (evaluate_model). Similarly, eval.py is reorganized to load the model and perform predictions efficiently."
+        ),
+    },
+    {
+        "Action": "Retrieve Model",
+        "Observation": "CNN and BiLSTM retrieved.",
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The model trained over the specified number of epochs. Training and validation loss values are recorded for each epoch, "
+            "the decrease in loss indicates improved model performance."
+        )
+    },
+    {
+        "Action": "Evaluation",
+        "Observation": predefined_observation,
+    }
+]
+def info_to_message(info):
+    msg = ""
+    for k, v in info.items():
+        if isinstance(v, dict):
+            tempv = v
+            v = ""
+            for k2, v2 in tempv.items():
+                v += f"{k2}:\n  {v2}\n"
+        v = User.indent_text(v, 2)
+        msg += '-' * 64
+        msg += '\n'
+        msg += f"{k}:\n{v}\n"
+    return msg
+
+
+def handle_example_click(example_index):
+    global index_ex
+    index_ex = example_index
+    return load_example(index_ex)  # Simply return the text to display it in the textbox
+
+# Gradio Interface
+with gr.Blocks(css=".gr-examples-label {display: none;}", theme=gr.themes.Default()) as app:
+    gr.Markdown("# MLR- Copilot: Machine Learning Research based on LLM Agents")
+    gr.Markdown("###   ")
+    gr.Markdown("## <span style='color:Orange;'> This UI is for predefined example demo only.</span>")
+    gr.Markdown("## <span style='color:Orange;'> To reproduce the results please use [Github](https://github.com/du-nlp-lab/MLR-Copilot/)</span>")
+
+
+
+    gr.Markdown("MLR-Copilot is a framework where LLMs mimic researchers’ thought processes, designed to enhance the productivity of machine learning research by automating the generation and implementation of research ideas. It begins with a research paper, autonomously generating and validating these ideas, while incorporating human feedback to help reach executable research outcomes.")
+
+
+    gr.Markdown("##    ⬇️  Click an example at bottom to start！")
+
+
+    # Use state variables to store generated hypothesis and experiment plan
+    hypothesis_state = gr.State("")
+    experiment_plan_state = gr.State("")
+
+    ########## Phase 1:  Research Idea Generation Tab ##############
+    with gr.Tab("💡Stage 1: Research Idea Generation"):
+        gr.Markdown("### Extract Research Elements and Generate Research Ideas")
+        with gr.Row():
+            with gr.Column():
+                paper_text_input = gr.Textbox(value="", lines=10, label="📑 Research Paper Text", interactive=False)
+
+                # extract_button = gr.Button("🔍 Extract Research Elements")
+                with gr.Row():
+                    tasks_output = gr.Textbox(placeholder="Research task definition", label="Research Tasks", lines=2, interactive=False)
+                    gaps_output = gr.Textbox(placeholder="Research gaps of current works", label="Research Gaps", lines=2, interactive=False)
+                    keywords_output = gr.Textbox(placeholder="Paper keywords", label="Keywords", lines=2, interactive=False)
+                    recent_works_output = gr.Textbox(placeholder="Recent works extracted from Semantic Scholar", label="Recent Works", lines=2, interactive=False)
+            with gr.Column():
+                with gr.Row():  # Move the button to the top 
+                    generate_button = gr.Button("✍️ Generate Research Hypothesis & Experiment Plan")
+                with gr.Group():
+                    gr.Markdown("### 🌟 Research Idea")
+                    with gr.Row():
+                        hypothesis_output = gr.Textbox(label="Generated Hypothesis", lines=20, interactive=False)
+                        experiment_plan_output = gr.Textbox(label="Generated Experiment Plan", lines=20, interactive=False)
+        
+    
+        # Step 1: Extract Research Elements
+        # extract_button.click(
+        #     fn=extract_research_elements, 
+        #     inputs=paper_text_input, 
+        #     outputs=[tasks_output, gaps_output, keywords_output, recent_works_output]
+        # )
+
+        generate_button.click(
+            fn=generate_and_store, 
+            inputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output], 
+            outputs=[hypothesis_output, experiment_plan_output, hypothesis_state, experiment_plan_state]
+        )
+
+        gr.Examples(
+            examples=example_text,
+            inputs=[paper_text_input],
+            outputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output],
+            fn=load_example_and_set_index,
+            run_on_click = True,
+            # label="⬇️ Click an example to load"
+        )
+        
+
+
+
+    ########## Phase 2 & 3:  Experiment implementation and execution ##############
+    with gr.Tab("🧪 Stage 2 & Stage 3: Experiment implementation and execution"):
+        gr.Markdown("### Interact with the ExperimentAgent")
+
+        with gr.Row():
+            with gr.Column():
+                with gr.Group():
+                    gr.Markdown("### 🌟 Generated Research Idea")
+                    with gr.Row():
+                        idea_input = gr.Textbox(label="Generated Research Hypothesis", lines=30, interactive=False)
+                        plan_input = gr.Textbox(label="Generated Experiment Plan", lines=30, interactive=False)
+        
+            with gr.Column():
+                start_exp_agnet = gr.Button("⚙️ Start / Reset ExperimentAgent", elem_classes=["agent-btn"])
+                with gr.Group():
+                    gr.Markdown("### Implementation + Execution Log")    
+                    log = gr.Textbox(label="📖 Execution Log", lines=20, interactive=False)
+                    code_display = gr.Code(label="🧑‍💻 Implementation", language="python", interactive=False)
+
+            with gr.Column():
+                response = gr.Textbox(label="🤖 ExperimentAgent Response", lines=30, interactive=False)
+                feedback = gr.Textbox(placeholder="N/A", label="🧑‍🔬 User Feedback", lines=3, interactive=True)
+                submit_button = gr.Button("Submit", elem_classes=["Submit-btn"])
+        
+        hypothesis_state.change(
+            fn=load_phase_2_inputs, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[idea_input, plan_input, code_display]
+        )
+        
+        # Start research agent
+        start_exp_agnet.click(
+            fn=start_experiment_agent, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[code_display, log, response, feedback]
+        )
+        
+        submit_button.click(
+            fn=submit_feedback,
+            inputs=[feedback, log, response],
+            outputs=[log, response, code_display, feedback]
+        )
+
+# Test
+if __name__ == "__main__":
+    step_index = 0
+    app.launch(share=True)

.history/app_20250404175236.py

@@ -0,0 +1,335 @@
+import gradio as gr
+from pathlib import Path
+from reactagent.environment import Environment
+from reactagent.agents.agent_research import ResearchAgent
+from reactagent.runner import create_parser
+from reactagent import llm
+from reactagent.users.user import User
+import os
+import json
+
+
+# Global variables to store session state
+env = None
+agent = None
+state_example = False
+state_extract = False
+state_generate = False
+state_agent = False
+state_complete = False
+index_ex = "1"
+
+example_text = [
+    "Research Paper 1: Dataset and Baseline for Automatic Student Feedback Analysis",
+    "Research Paper 2: An Empirical Study on the Impact of Code Review on Software Quality"
+]
+
+
+# Load example JSON file
+def load_example_data():
+    with open("example/example_data.json", "r") as json_file:
+        example_data = json.load(json_file)
+    
+    for idx in example_data.keys():
+        try:
+            file = example_data[idx]["code_init"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_init"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")
+        try:
+            file = example_data[idx]["code_final"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_final"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")     
+    return example_data
+
+example_data = load_example_data()
+
+# Function to handle the selection of an example and populate the respective fields
+def load_example(example_id):
+    global index_ex 
+    index_ex = str(example_id)
+    example = example_data[index_ex]
+    paper_text = 'Title:\t' +  example['title'] + '\n\nAbstract:\t' + example['abstract']
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return [paper_text, tasks, gaps, keywords, recent_works]
+
+example_text = [load_example(1)[0], load_example(2)[0]]
+
+
+# Function to handle example clicks
+def load_example_and_set_index(paper_text_input):
+    global index_ex, state_example
+    state_example = True
+    index_ex = str(example_text.index(paper_text_input) + 1)
+    example =  load_example(index_ex)
+
+    return example
+
+
+
+########## Phase 1 ##############
+
+def extract_research_elements(paper_text):
+    global state_extract, index_ex, state_example
+    if not state_example or paper_text == "":
+        return "", "", "", ""
+    state_extract = True
+    if not paper_text.strip().startswith("Title:\t" + example_data[index_ex]["title"]):
+        print("Mismatch detected.")
+        print(paper_text)
+        return "", "", "", ""
+    example = example_data[index_ex]
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return tasks, gaps, keywords, recent_works
+
+
+# Step 2: Generate Research Hypothesis and Experiment Plan
+def generate_and_store(paper_text, tasks, gaps, keywords, recent_works):
+    if (not state_example):
+        return "", "", "", ""
+    global state_generate, index_ex
+    state_generate = True
+    hypothesis = example_data[index_ex]['hypothesis']
+    experiment_plan = example_data[index_ex]['experiment_plan']
+    return hypothesis, experiment_plan, hypothesis, experiment_plan
+
+##########  Phase 2 & 3  ##############
+def start_experiment_agent(hypothesis, plan):
+    if (not state_extract or not state_generate or not state_example):
+        return "", "", ""
+    global state_agent, step_index, state_complete
+    state_agent = True
+    step_index = 0
+    state_complete = False
+    # predefined_message = f"Implement the following hypothesis and experiment plan:\n\nHypothesis:\n{hypothesis}\n\nExperiment Plan:\n{plan}"
+    return example_data[index_ex]['code_init'], predefined_action_log, "", ""
+
+def submit_feedback(user_feedback, history, previous_response):
+    if (not state_extract or not  state_generate or not state_agent or not state_example):
+        return "", "", ""
+    global step_index, state_complete
+    step_index += 1
+    msg = history
+    if step_index < len(process_steps):
+        msg += previous_response + "\nUser feedback:" + user_feedback + "\n\n"
+        response_info = process_steps[step_index]
+        response = info_to_message(response_info)  # Convert dictionary to formatted string
+        response += "Please provide feedback based on the history, response entries, and observation, and questions: "
+        step_index += 1
+        msg += response
+    else:
+        state_complete = True
+        response = "Agent Finished."
+        
+    return msg, response, example_data[index_ex]['code_init'] if state_complete else example_data[index_ex]['code_final'], ""
+
+def load_phase_2_inputs(hypothesis, plan):
+    return hypothesis, plan, "# Code implementation will be displayed here after Start ExperimentAgent."
+
+
+
+predefined_action_log = """
+[Reasoning]: To understand the initial structure and functionality of train.py for effective improvements.
+[Action]: Inspect Script (train.py)
+Input: {"script_name": "train.py", "start_line_number": "1", "end_line_number": "74"}
+Objective: Understand the training script, including data processing, [...]
+[Observation]: The train.py script imports [...]. Sets random seeds [...]. Defines [...] Placeholder functions [...] exist without implementation. [...]
+[Feedback]: The script structure is clear, but key functions (train_model, predict) need proper implementation for proposed model training and prediction.\n
+"""
+
+
+predefined_observation = """
+Epoch [1/10],
+Train MSE: 0.543,
+Test MSE: 0.688
+Epoch [2/10],
+Train MSE: 0.242,
+Test MSE: 0.493\n
+"""
+
+# Initialize the global step_index and history
+process_steps = [
+    {
+        "Action": "Inspect Script Lines (train.py)",
+        "Observation": (
+            "The train.py script imports necessary libraries (e.g., pandas, sklearn, torch). "
+            "Sets random seeds for reproducibility. Defines compute_metrics_for_regression function "
+            "to calculate RMSE for different dimensions. Placeholder functions train_model and "
+            "predict exist without implementations."
+        ),
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The script executed successfully. Generated embeddings using the BERT model. Completed "
+            "the training process without errors. Metrics calculation placeholders indicated areas needing implementation."
+        ),
+    },
+    {
+        "Action": "Edit Script (train.py)",
+        "Observation": (
+            "Edited train.py to separate data loading, model definition, training loop, and evaluation into distinct functions. "
+            "The edited train.py now has clearly defined functions"
+            "for data loading (load_data), model definition (build_model), "
+            "training (train_model), and evaluation (evaluate_model). Similarly, eval.py is reorganized to load the model and perform predictions efficiently."
+        ),
+    },
+    {
+        "Action": "Retrieve Model",
+        "Observation": "CNN and BiLSTM retrieved.",
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The model trained over the specified number of epochs. Training and validation loss values are recorded for each epoch, "
+            "the decrease in loss indicates improved model performance."
+        )
+    },
+    {
+        "Action": "Evaluation",
+        "Observation": predefined_observation,
+    }
+]
+def info_to_message(info):
+    msg = ""
+    for k, v in info.items():
+        if isinstance(v, dict):
+            tempv = v
+            v = ""
+            for k2, v2 in tempv.items():
+                v += f"{k2}:\n  {v2}\n"
+        v = User.indent_text(v, 2)
+        msg += '-' * 64
+        msg += '\n'
+        msg += f"{k}:\n{v}\n"
+    return msg
+
+
+def handle_example_click(example_index):
+    global index_ex
+    index_ex = example_index
+    return load_example(index_ex)  # Simply return the text to display it in the textbox
+
+# Gradio Interface
+with gr.Blocks(css=".gr-examples-label {display: none;}", theme=gr.themes.Default()) as app:
+    gr.Markdown("# MLR- Copilot: Machine Learning Research based on LLM Agents")
+    gr.Markdown("###   ")
+    gr.Markdown("## <span style='color:Orange;'> This UI is for predefined example demo only.</span>")
+    gr.Markdown("## <span style='color:Orange;'> To reproduce the results please use [Github](https://github.com/du-nlp-lab/MLR-Copilot/)</span>")
+
+
+
+    gr.Markdown("MLR-Copilot is a framework where LLMs mimic researchers’ thought processes, designed to enhance the productivity of machine learning research by automating the generation and implementation of research ideas. It begins with a research paper, autonomously generating and validating these ideas, while incorporating human feedback to help reach executable research outcomes.")
+
+
+    gr.Markdown("##    ⬇️  Click an example at bottom to start✅")
+
+
+    # Use state variables to store generated hypothesis and experiment plan
+    hypothesis_state = gr.State("")
+    experiment_plan_state = gr.State("")
+
+    ########## Phase 1:  Research Idea Generation Tab ##############
+    with gr.Tab("💡Stage 1: Research Idea Generation"):
+        gr.Markdown("### Extract Research Elements and Generate Research Ideas")
+        with gr.Row():
+            with gr.Column():
+                paper_text_input = gr.Textbox(value="", lines=10, label="📑 Research Paper Text", interactive=False)
+
+                # extract_button = gr.Button("🔍 Extract Research Elements")
+                with gr.Row():
+                    tasks_output = gr.Textbox(placeholder="Research task definition", label="Research Tasks", lines=2, interactive=False)
+                    gaps_output = gr.Textbox(placeholder="Research gaps of current works", label="Research Gaps", lines=2, interactive=False)
+                    keywords_output = gr.Textbox(placeholder="Paper keywords", label="Keywords", lines=2, interactive=False)
+                    recent_works_output = gr.Textbox(placeholder="Recent works extracted from Semantic Scholar", label="Recent Works", lines=2, interactive=False)
+            with gr.Column():
+                with gr.Row():  # Move the button to the top 
+                    generate_button = gr.Button("✍️ Generate Research Hypothesis & Experiment Plan")
+                with gr.Group():
+                    gr.Markdown("### 🌟 Research Idea")
+                    with gr.Row():
+                        hypothesis_output = gr.Textbox(label="Generated Hypothesis", lines=20, interactive=False)
+                        experiment_plan_output = gr.Textbox(label="Generated Experiment Plan", lines=20, interactive=False)
+        
+    
+        # Step 1: Extract Research Elements
+        # extract_button.click(
+        #     fn=extract_research_elements, 
+        #     inputs=paper_text_input, 
+        #     outputs=[tasks_output, gaps_output, keywords_output, recent_works_output]
+        # )
+
+        generate_button.click(
+            fn=generate_and_store, 
+            inputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output], 
+            outputs=[hypothesis_output, experiment_plan_output, hypothesis_state, experiment_plan_state]
+        )
+
+        gr.Examples(
+            examples=example_text,
+            inputs=[paper_text_input],
+            outputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output],
+            fn=load_example_and_set_index,
+            run_on_click = True,
+            # label="⬇️ Click an example to load"
+        )
+        
+
+
+
+    ########## Phase 2 & 3:  Experiment implementation and execution ##############
+    with gr.Tab("🧪 Stage 2 & Stage 3: Experiment implementation and execution"):
+        gr.Markdown("### Interact with the ExperimentAgent")
+
+        with gr.Row():
+            with gr.Column():
+                with gr.Group():
+                    gr.Markdown("### 🌟 Generated Research Idea")
+                    with gr.Row():
+                        idea_input = gr.Textbox(label="Generated Research Hypothesis", lines=30, interactive=False)
+                        plan_input = gr.Textbox(label="Generated Experiment Plan", lines=30, interactive=False)
+        
+            with gr.Column():
+                start_exp_agnet = gr.Button("⚙️ Start / Reset ExperimentAgent", elem_classes=["agent-btn"])
+                with gr.Group():
+                    gr.Markdown("### Implementation + Execution Log")    
+                    log = gr.Textbox(label="📖 Execution Log", lines=20, interactive=False)
+                    code_display = gr.Code(label="🧑‍💻 Implementation", language="python", interactive=False)
+
+            with gr.Column():
+                response = gr.Textbox(label="🤖 ExperimentAgent Response", lines=30, interactive=False)
+                feedback = gr.Textbox(placeholder="N/A", label="🧑‍🔬 User Feedback", lines=3, interactive=True)
+                submit_button = gr.Button("Submit", elem_classes=["Submit-btn"])
+        
+        hypothesis_state.change(
+            fn=load_phase_2_inputs, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[idea_input, plan_input, code_display]
+        )
+        
+        # Start research agent
+        start_exp_agnet.click(
+            fn=start_experiment_agent, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[code_display, log, response, feedback]
+        )
+        
+        submit_button.click(
+            fn=submit_feedback,
+            inputs=[feedback, log, response],
+            outputs=[log, response, code_display, feedback]
+        )
+
+# Test
+if __name__ == "__main__":
+    step_index = 0
+    app.launch(share=True)

.history/app_20250404175239.py

@@ -0,0 +1,335 @@
+import gradio as gr
+from pathlib import Path
+from reactagent.environment import Environment
+from reactagent.agents.agent_research import ResearchAgent
+from reactagent.runner import create_parser
+from reactagent import llm
+from reactagent.users.user import User
+import os
+import json
+
+
+# Global variables to store session state
+env = None
+agent = None
+state_example = False
+state_extract = False
+state_generate = False
+state_agent = False
+state_complete = False
+index_ex = "1"
+
+example_text = [
+    "Research Paper 1: Dataset and Baseline for Automatic Student Feedback Analysis",
+    "Research Paper 2: An Empirical Study on the Impact of Code Review on Software Quality"
+]
+
+
+# Load example JSON file
+def load_example_data():
+    with open("example/example_data.json", "r") as json_file:
+        example_data = json.load(json_file)
+    
+    for idx in example_data.keys():
+        try:
+            file = example_data[idx]["code_init"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_init"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")
+        try:
+            file = example_data[idx]["code_final"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_final"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")     
+    return example_data
+
+example_data = load_example_data()
+
+# Function to handle the selection of an example and populate the respective fields
+def load_example(example_id):
+    global index_ex 
+    index_ex = str(example_id)
+    example = example_data[index_ex]
+    paper_text = 'Title:\t' +  example['title'] + '\n\nAbstract:\t' + example['abstract']
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return [paper_text, tasks, gaps, keywords, recent_works]
+
+example_text = [load_example(1)[0], load_example(2)[0]]
+
+
+# Function to handle example clicks
+def load_example_and_set_index(paper_text_input):
+    global index_ex, state_example
+    state_example = True
+    index_ex = str(example_text.index(paper_text_input) + 1)
+    example =  load_example(index_ex)
+
+    return example
+
+
+
+########## Phase 1 ##############
+
+def extract_research_elements(paper_text):
+    global state_extract, index_ex, state_example
+    if not state_example or paper_text == "":
+        return "", "", "", ""
+    state_extract = True
+    if not paper_text.strip().startswith("Title:\t" + example_data[index_ex]["title"]):
+        print("Mismatch detected.")
+        print(paper_text)
+        return "", "", "", ""
+    example = example_data[index_ex]
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return tasks, gaps, keywords, recent_works
+
+
+# Step 2: Generate Research Hypothesis and Experiment Plan
+def generate_and_store(paper_text, tasks, gaps, keywords, recent_works):
+    if (not state_example):
+        return "", "", "", ""
+    global state_generate, index_ex
+    state_generate = True
+    hypothesis = example_data[index_ex]['hypothesis']
+    experiment_plan = example_data[index_ex]['experiment_plan']
+    return hypothesis, experiment_plan, hypothesis, experiment_plan
+
+##########  Phase 2 & 3  ##############
+def start_experiment_agent(hypothesis, plan):
+    if (not state_extract or not state_generate or not state_example):
+        return "", "", ""
+    global state_agent, step_index, state_complete
+    state_agent = True
+    step_index = 0
+    state_complete = False
+    # predefined_message = f"Implement the following hypothesis and experiment plan:\n\nHypothesis:\n{hypothesis}\n\nExperiment Plan:\n{plan}"
+    return example_data[index_ex]['code_init'], predefined_action_log, "", ""
+
+def submit_feedback(user_feedback, history, previous_response):
+    if (not state_extract or not  state_generate or not state_agent or not state_example):
+        return "", "", ""
+    global step_index, state_complete
+    step_index += 1
+    msg = history
+    if step_index < len(process_steps):
+        msg += previous_response + "\nUser feedback:" + user_feedback + "\n\n"
+        response_info = process_steps[step_index]
+        response = info_to_message(response_info)  # Convert dictionary to formatted string
+        response += "Please provide feedback based on the history, response entries, and observation, and questions: "
+        step_index += 1
+        msg += response
+    else:
+        state_complete = True
+        response = "Agent Finished."
+        
+    return msg, response, example_data[index_ex]['code_init'] if state_complete else example_data[index_ex]['code_final'], ""
+
+def load_phase_2_inputs(hypothesis, plan):
+    return hypothesis, plan, "# Code implementation will be displayed here after Start ExperimentAgent."
+
+
+
+predefined_action_log = """
+[Reasoning]: To understand the initial structure and functionality of train.py for effective improvements.
+[Action]: Inspect Script (train.py)
+Input: {"script_name": "train.py", "start_line_number": "1", "end_line_number": "74"}
+Objective: Understand the training script, including data processing, [...]
+[Observation]: The train.py script imports [...]. Sets random seeds [...]. Defines [...] Placeholder functions [...] exist without implementation. [...]
+[Feedback]: The script structure is clear, but key functions (train_model, predict) need proper implementation for proposed model training and prediction.\n
+"""
+
+
+predefined_observation = """
+Epoch [1/10],
+Train MSE: 0.543,
+Test MSE: 0.688
+Epoch [2/10],
+Train MSE: 0.242,
+Test MSE: 0.493\n
+"""
+
+# Initialize the global step_index and history
+process_steps = [
+    {
+        "Action": "Inspect Script Lines (train.py)",
+        "Observation": (
+            "The train.py script imports necessary libraries (e.g., pandas, sklearn, torch). "
+            "Sets random seeds for reproducibility. Defines compute_metrics_for_regression function "
+            "to calculate RMSE for different dimensions. Placeholder functions train_model and "
+            "predict exist without implementations."
+        ),
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The script executed successfully. Generated embeddings using the BERT model. Completed "
+            "the training process without errors. Metrics calculation placeholders indicated areas needing implementation."
+        ),
+    },
+    {
+        "Action": "Edit Script (train.py)",
+        "Observation": (
+            "Edited train.py to separate data loading, model definition, training loop, and evaluation into distinct functions. "
+            "The edited train.py now has clearly defined functions"
+            "for data loading (load_data), model definition (build_model), "
+            "training (train_model), and evaluation (evaluate_model). Similarly, eval.py is reorganized to load the model and perform predictions efficiently."
+        ),
+    },
+    {
+        "Action": "Retrieve Model",
+        "Observation": "CNN and BiLSTM retrieved.",
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The model trained over the specified number of epochs. Training and validation loss values are recorded for each epoch, "
+            "the decrease in loss indicates improved model performance."
+        )
+    },
+    {
+        "Action": "Evaluation",
+        "Observation": predefined_observation,
+    }
+]
+def info_to_message(info):
+    msg = ""
+    for k, v in info.items():
+        if isinstance(v, dict):
+            tempv = v
+            v = ""
+            for k2, v2 in tempv.items():
+                v += f"{k2}:\n  {v2}\n"
+        v = User.indent_text(v, 2)
+        msg += '-' * 64
+        msg += '\n'
+        msg += f"{k}:\n{v}\n"
+    return msg
+
+
+def handle_example_click(example_index):
+    global index_ex
+    index_ex = example_index
+    return load_example(index_ex)  # Simply return the text to display it in the textbox
+
+# Gradio Interface
+with gr.Blocks(css=".gr-examples-label {display: none;}", theme=gr.themes.Default()) as app:
+    gr.Markdown("# MLR- Copilot: Machine Learning Research based on LLM Agents")
+    gr.Markdown("###   ")
+    gr.Markdown("## <span style='color:Orange;'> This UI is for predefined example demo only.</span>")
+    gr.Markdown("## <span style='color:Orange;'> To reproduce the results please use [Github](https://github.com/du-nlp-lab/MLR-Copilot/)</span>")
+
+
+
+    gr.Markdown("MLR-Copilot is a framework where LLMs mimic researchers’ thought processes, designed to enhance the productivity of machine learning research by automating the generation and implementation of research ideas. It begins with a research paper, autonomously generating and validating these ideas, while incorporating human feedback to help reach executable research outcomes.")
+
+
+    gr.Markdown("##    ⬇️  Click an example at bottom to start ✅")
+
+
+    # Use state variables to store generated hypothesis and experiment plan
+    hypothesis_state = gr.State("")
+    experiment_plan_state = gr.State("")
+
+    ########## Phase 1:  Research Idea Generation Tab ##############
+    with gr.Tab("💡Stage 1: Research Idea Generation"):
+        gr.Markdown("### Extract Research Elements and Generate Research Ideas")
+        with gr.Row():
+            with gr.Column():
+                paper_text_input = gr.Textbox(value="", lines=10, label="📑 Research Paper Text", interactive=False)
+
+                # extract_button = gr.Button("🔍 Extract Research Elements")
+                with gr.Row():
+                    tasks_output = gr.Textbox(placeholder="Research task definition", label="Research Tasks", lines=2, interactive=False)
+                    gaps_output = gr.Textbox(placeholder="Research gaps of current works", label="Research Gaps", lines=2, interactive=False)
+                    keywords_output = gr.Textbox(placeholder="Paper keywords", label="Keywords", lines=2, interactive=False)
+                    recent_works_output = gr.Textbox(placeholder="Recent works extracted from Semantic Scholar", label="Recent Works", lines=2, interactive=False)
+            with gr.Column():
+                with gr.Row():  # Move the button to the top 
+                    generate_button = gr.Button("✍️ Generate Research Hypothesis & Experiment Plan")
+                with gr.Group():
+                    gr.Markdown("### 🌟 Research Idea")
+                    with gr.Row():
+                        hypothesis_output = gr.Textbox(label="Generated Hypothesis", lines=20, interactive=False)
+                        experiment_plan_output = gr.Textbox(label="Generated Experiment Plan", lines=20, interactive=False)
+        
+    
+        # Step 1: Extract Research Elements
+        # extract_button.click(
+        #     fn=extract_research_elements, 
+        #     inputs=paper_text_input, 
+        #     outputs=[tasks_output, gaps_output, keywords_output, recent_works_output]
+        # )
+
+        generate_button.click(
+            fn=generate_and_store, 
+            inputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output], 
+            outputs=[hypothesis_output, experiment_plan_output, hypothesis_state, experiment_plan_state]
+        )
+
+        gr.Examples(
+            examples=example_text,
+            inputs=[paper_text_input],
+            outputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output],
+            fn=load_example_and_set_index,
+            run_on_click = True,
+            # label="⬇️ Click an example to load"
+        )
+        
+
+
+
+    ########## Phase 2 & 3:  Experiment implementation and execution ##############
+    with gr.Tab("🧪 Stage 2 & Stage 3: Experiment implementation and execution"):
+        gr.Markdown("### Interact with the ExperimentAgent")
+
+        with gr.Row():
+            with gr.Column():
+                with gr.Group():
+                    gr.Markdown("### 🌟 Generated Research Idea")
+                    with gr.Row():
+                        idea_input = gr.Textbox(label="Generated Research Hypothesis", lines=30, interactive=False)
+                        plan_input = gr.Textbox(label="Generated Experiment Plan", lines=30, interactive=False)
+        
+            with gr.Column():
+                start_exp_agnet = gr.Button("⚙️ Start / Reset ExperimentAgent", elem_classes=["agent-btn"])
+                with gr.Group():
+                    gr.Markdown("### Implementation + Execution Log")    
+                    log = gr.Textbox(label="📖 Execution Log", lines=20, interactive=False)
+                    code_display = gr.Code(label="🧑‍💻 Implementation", language="python", interactive=False)
+
+            with gr.Column():
+                response = gr.Textbox(label="🤖 ExperimentAgent Response", lines=30, interactive=False)
+                feedback = gr.Textbox(placeholder="N/A", label="🧑‍🔬 User Feedback", lines=3, interactive=True)
+                submit_button = gr.Button("Submit", elem_classes=["Submit-btn"])
+        
+        hypothesis_state.change(
+            fn=load_phase_2_inputs, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[idea_input, plan_input, code_display]
+        )
+        
+        # Start research agent
+        start_exp_agnet.click(
+            fn=start_experiment_agent, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[code_display, log, response, feedback]
+        )
+        
+        submit_button.click(
+            fn=submit_feedback,
+            inputs=[feedback, log, response],
+            outputs=[log, response, code_display, feedback]
+        )
+
+# Test
+if __name__ == "__main__":
+    step_index = 0
+    app.launch(share=True)

.history/app_20250404175250.py

@@ -0,0 +1,335 @@
+import gradio as gr
+from pathlib import Path
+from reactagent.environment import Environment
+from reactagent.agents.agent_research import ResearchAgent
+from reactagent.runner import create_parser
+from reactagent import llm
+from reactagent.users.user import User
+import os
+import json
+
+
+# Global variables to store session state
+env = None
+agent = None
+state_example = False
+state_extract = False
+state_generate = False
+state_agent = False
+state_complete = False
+index_ex = "1"
+
+example_text = [
+    "Research Paper 1: Dataset and Baseline for Automatic Student Feedback Analysis",
+    "Research Paper 2: An Empirical Study on the Impact of Code Review on Software Quality"
+]
+
+
+# Load example JSON file
+def load_example_data():
+    with open("example/example_data.json", "r") as json_file:
+        example_data = json.load(json_file)
+    
+    for idx in example_data.keys():
+        try:
+            file = example_data[idx]["code_init"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_init"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")
+        try:
+            file = example_data[idx]["code_final"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_final"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")     
+    return example_data
+
+example_data = load_example_data()
+
+# Function to handle the selection of an example and populate the respective fields
+def load_example(example_id):
+    global index_ex 
+    index_ex = str(example_id)
+    example = example_data[index_ex]
+    paper_text = 'Title:\t' +  example['title'] + '\n\nAbstract:\t' + example['abstract']
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return [paper_text, tasks, gaps, keywords, recent_works]
+
+example_text = [load_example(1)[0], load_example(2)[0]]
+
+
+# Function to handle example clicks
+def load_example_and_set_index(paper_text_input):
+    global index_ex, state_example
+    state_example = True
+    index_ex = str(example_text.index(paper_text_input) + 1)
+    example =  load_example(index_ex)
+
+    return example
+
+
+
+########## Phase 1 ##############
+
+def extract_research_elements(paper_text):
+    global state_extract, index_ex, state_example
+    if not state_example or paper_text == "":
+        return "", "", "", ""
+    state_extract = True
+    if not paper_text.strip().startswith("Title:\t" + example_data[index_ex]["title"]):
+        print("Mismatch detected.")
+        print(paper_text)
+        return "", "", "", ""
+    example = example_data[index_ex]
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return tasks, gaps, keywords, recent_works
+
+
+# Step 2: Generate Research Hypothesis and Experiment Plan
+def generate_and_store(paper_text, tasks, gaps, keywords, recent_works):
+    if (not state_example):
+        return "", "", "", ""
+    global state_generate, index_ex
+    state_generate = True
+    hypothesis = example_data[index_ex]['hypothesis']
+    experiment_plan = example_data[index_ex]['experiment_plan']
+    return hypothesis, experiment_plan, hypothesis, experiment_plan
+
+##########  Phase 2 & 3  ##############
+def start_experiment_agent(hypothesis, plan):
+    if (not state_extract or not state_generate or not state_example):
+        return "", "", ""
+    global state_agent, step_index, state_complete
+    state_agent = True
+    step_index = 0
+    state_complete = False
+    # predefined_message = f"Implement the following hypothesis and experiment plan:\n\nHypothesis:\n{hypothesis}\n\nExperiment Plan:\n{plan}"
+    return example_data[index_ex]['code_init'], predefined_action_log, "", ""
+
+def submit_feedback(user_feedback, history, previous_response):
+    if (not state_extract or not  state_generate or not state_agent or not state_example):
+        return "", "", ""
+    global step_index, state_complete
+    step_index += 1
+    msg = history
+    if step_index < len(process_steps):
+        msg += previous_response + "\nUser feedback:" + user_feedback + "\n\n"
+        response_info = process_steps[step_index]
+        response = info_to_message(response_info)  # Convert dictionary to formatted string
+        response += "Please provide feedback based on the history, response entries, and observation, and questions: "
+        step_index += 1
+        msg += response
+    else:
+        state_complete = True
+        response = "Agent Finished."
+        
+    return msg, response, example_data[index_ex]['code_init'] if state_complete else example_data[index_ex]['code_final'], ""
+
+def load_phase_2_inputs(hypothesis, plan):
+    return hypothesis, plan, "# Code implementation will be displayed here after Start ExperimentAgent."
+
+
+
+predefined_action_log = """
+[Reasoning]: To understand the initial structure and functionality of train.py for effective improvements.
+[Action]: Inspect Script (train.py)
+Input: {"script_name": "train.py", "start_line_number": "1", "end_line_number": "74"}
+Objective: Understand the training script, including data processing, [...]
+[Observation]: The train.py script imports [...]. Sets random seeds [...]. Defines [...] Placeholder functions [...] exist without implementation. [...]
+[Feedback]: The script structure is clear, but key functions (train_model, predict) need proper implementation for proposed model training and prediction.\n
+"""
+
+
+predefined_observation = """
+Epoch [1/10],
+Train MSE: 0.543,
+Test MSE: 0.688
+Epoch [2/10],
+Train MSE: 0.242,
+Test MSE: 0.493\n
+"""
+
+# Initialize the global step_index and history
+process_steps = [
+    {
+        "Action": "Inspect Script Lines (train.py)",
+        "Observation": (
+            "The train.py script imports necessary libraries (e.g., pandas, sklearn, torch). "
+            "Sets random seeds for reproducibility. Defines compute_metrics_for_regression function "
+            "to calculate RMSE for different dimensions. Placeholder functions train_model and "
+            "predict exist without implementations."
+        ),
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The script executed successfully. Generated embeddings using the BERT model. Completed "
+            "the training process without errors. Metrics calculation placeholders indicated areas needing implementation."
+        ),
+    },
+    {
+        "Action": "Edit Script (train.py)",
+        "Observation": (
+            "Edited train.py to separate data loading, model definition, training loop, and evaluation into distinct functions. "
+            "The edited train.py now has clearly defined functions"
+            "for data loading (load_data), model definition (build_model), "
+            "training (train_model), and evaluation (evaluate_model). Similarly, eval.py is reorganized to load the model and perform predictions efficiently."
+        ),
+    },
+    {
+        "Action": "Retrieve Model",
+        "Observation": "CNN and BiLSTM retrieved.",
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The model trained over the specified number of epochs. Training and validation loss values are recorded for each epoch, "
+            "the decrease in loss indicates improved model performance."
+        )
+    },
+    {
+        "Action": "Evaluation",
+        "Observation": predefined_observation,
+    }
+]
+def info_to_message(info):
+    msg = ""
+    for k, v in info.items():
+        if isinstance(v, dict):
+            tempv = v
+            v = ""
+            for k2, v2 in tempv.items():
+                v += f"{k2}:\n  {v2}\n"
+        v = User.indent_text(v, 2)
+        msg += '-' * 64
+        msg += '\n'
+        msg += f"{k}:\n{v}\n"
+    return msg
+
+
+def handle_example_click(example_index):
+    global index_ex
+    index_ex = example_index
+    return load_example(index_ex)  # Simply return the text to display it in the textbox
+
+# Gradio Interface
+with gr.Blocks(css=".gr-examples-label {display: none;}", theme=gr.themes.Default()) as app:
+    gr.Markdown("# MLR- Copilot: Machine Learning Research based on LLM Agents")
+    gr.Markdown("###   ")
+    gr.Markdown("## <span style='color:Orange;'> This UI is for predefined example demo only.</span>")
+    gr.Markdown("## <span style='color:Orange;'> To reproduce the results please use [Github](https://github.com/du-nlp-lab/MLR-Copilot/)</span>")
+
+
+
+    gr.Markdown("MLR-Copilot is a framework where LLMs mimic researchers’ thought processes, designed to enhance the productivity of machine learning research by automating the generation and implementation of research ideas. It begins with a research paper, autonomously generating and validating these ideas, while incorporating human feedback to help reach executable research outcomes.")
+
+
+    gr.Markdown("##     ✅  Click an example at bottom to start  ⬇️")
+
+
+    # Use state variables to store generated hypothesis and experiment plan
+    hypothesis_state = gr.State("")
+    experiment_plan_state = gr.State("")
+
+    ########## Phase 1:  Research Idea Generation Tab ##############
+    with gr.Tab("💡Stage 1: Research Idea Generation"):
+        gr.Markdown("### Extract Research Elements and Generate Research Ideas")
+        with gr.Row():
+            with gr.Column():
+                paper_text_input = gr.Textbox(value="", lines=10, label="📑 Research Paper Text", interactive=False)
+
+                # extract_button = gr.Button("🔍 Extract Research Elements")
+                with gr.Row():
+                    tasks_output = gr.Textbox(placeholder="Research task definition", label="Research Tasks", lines=2, interactive=False)
+                    gaps_output = gr.Textbox(placeholder="Research gaps of current works", label="Research Gaps", lines=2, interactive=False)
+                    keywords_output = gr.Textbox(placeholder="Paper keywords", label="Keywords", lines=2, interactive=False)
+                    recent_works_output = gr.Textbox(placeholder="Recent works extracted from Semantic Scholar", label="Recent Works", lines=2, interactive=False)
+            with gr.Column():
+                with gr.Row():  # Move the button to the top 
+                    generate_button = gr.Button("✍️ Generate Research Hypothesis & Experiment Plan")
+                with gr.Group():
+                    gr.Markdown("### 🌟 Research Idea")
+                    with gr.Row():
+                        hypothesis_output = gr.Textbox(label="Generated Hypothesis", lines=20, interactive=False)
+                        experiment_plan_output = gr.Textbox(label="Generated Experiment Plan", lines=20, interactive=False)
+        
+    
+        # Step 1: Extract Research Elements
+        # extract_button.click(
+        #     fn=extract_research_elements, 
+        #     inputs=paper_text_input, 
+        #     outputs=[tasks_output, gaps_output, keywords_output, recent_works_output]
+        # )
+
+        generate_button.click(
+            fn=generate_and_store, 
+            inputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output], 
+            outputs=[hypothesis_output, experiment_plan_output, hypothesis_state, experiment_plan_state]
+        )
+
+        gr.Examples(
+            examples=example_text,
+            inputs=[paper_text_input],
+            outputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output],
+            fn=load_example_and_set_index,
+            run_on_click = True,
+            # label="⬇️ Click an example to load"
+        )
+        
+
+
+
+    ########## Phase 2 & 3:  Experiment implementation and execution ##############
+    with gr.Tab("🧪 Stage 2 & Stage 3: Experiment implementation and execution"):
+        gr.Markdown("### Interact with the ExperimentAgent")
+
+        with gr.Row():
+            with gr.Column():
+                with gr.Group():
+                    gr.Markdown("### 🌟 Generated Research Idea")
+                    with gr.Row():
+                        idea_input = gr.Textbox(label="Generated Research Hypothesis", lines=30, interactive=False)
+                        plan_input = gr.Textbox(label="Generated Experiment Plan", lines=30, interactive=False)
+        
+            with gr.Column():
+                start_exp_agnet = gr.Button("⚙️ Start / Reset ExperimentAgent", elem_classes=["agent-btn"])
+                with gr.Group():
+                    gr.Markdown("### Implementation + Execution Log")    
+                    log = gr.Textbox(label="📖 Execution Log", lines=20, interactive=False)
+                    code_display = gr.Code(label="🧑‍💻 Implementation", language="python", interactive=False)
+
+            with gr.Column():
+                response = gr.Textbox(label="🤖 ExperimentAgent Response", lines=30, interactive=False)
+                feedback = gr.Textbox(placeholder="N/A", label="🧑‍🔬 User Feedback", lines=3, interactive=True)
+                submit_button = gr.Button("Submit", elem_classes=["Submit-btn"])
+        
+        hypothesis_state.change(
+            fn=load_phase_2_inputs, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[idea_input, plan_input, code_display]
+        )
+        
+        # Start research agent
+        start_exp_agnet.click(
+            fn=start_experiment_agent, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[code_display, log, response, feedback]
+        )
+        
+        submit_button.click(
+            fn=submit_feedback,
+            inputs=[feedback, log, response],
+            outputs=[log, response, code_display, feedback]
+        )
+
+# Test
+if __name__ == "__main__":
+    step_index = 0
+    app.launch(share=True)

.history/app_20250404175251.py

@@ -0,0 +1,335 @@
+import gradio as gr
+from pathlib import Path
+from reactagent.environment import Environment
+from reactagent.agents.agent_research import ResearchAgent
+from reactagent.runner import create_parser
+from reactagent import llm
+from reactagent.users.user import User
+import os
+import json
+
+
+# Global variables to store session state
+env = None
+agent = None
+state_example = False
+state_extract = False
+state_generate = False
+state_agent = False
+state_complete = False
+index_ex = "1"
+
+example_text = [
+    "Research Paper 1: Dataset and Baseline for Automatic Student Feedback Analysis",
+    "Research Paper 2: An Empirical Study on the Impact of Code Review on Software Quality"
+]
+
+
+# Load example JSON file
+def load_example_data():
+    with open("example/example_data.json", "r") as json_file:
+        example_data = json.load(json_file)
+    
+    for idx in example_data.keys():
+        try:
+            file = example_data[idx]["code_init"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_init"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")
+        try:
+            file = example_data[idx]["code_final"]
+            with open(os.path.join("example", file), "r") as f:
+                example_data[idx]["code_final"] = f.read()
+        except FileNotFoundError:
+            print(f"File not found: {file}. Skipping key: {idx}")     
+    return example_data
+
+example_data = load_example_data()
+
+# Function to handle the selection of an example and populate the respective fields
+def load_example(example_id):
+    global index_ex 
+    index_ex = str(example_id)
+    example = example_data[index_ex]
+    paper_text = 'Title:\t' +  example['title'] + '\n\nAbstract:\t' + example['abstract']
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return [paper_text, tasks, gaps, keywords, recent_works]
+
+example_text = [load_example(1)[0], load_example(2)[0]]
+
+
+# Function to handle example clicks
+def load_example_and_set_index(paper_text_input):
+    global index_ex, state_example
+    state_example = True
+    index_ex = str(example_text.index(paper_text_input) + 1)
+    example =  load_example(index_ex)
+
+    return example
+
+
+
+########## Phase 1 ##############
+
+def extract_research_elements(paper_text):
+    global state_extract, index_ex, state_example
+    if not state_example or paper_text == "":
+        return "", "", "", ""
+    state_extract = True
+    if not paper_text.strip().startswith("Title:\t" + example_data[index_ex]["title"]):
+        print("Mismatch detected.")
+        print(paper_text)
+        return "", "", "", ""
+    example = example_data[index_ex]
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return tasks, gaps, keywords, recent_works
+
+
+# Step 2: Generate Research Hypothesis and Experiment Plan
+def generate_and_store(paper_text, tasks, gaps, keywords, recent_works):
+    if (not state_example):
+        return "", "", "", ""
+    global state_generate, index_ex
+    state_generate = True
+    hypothesis = example_data[index_ex]['hypothesis']
+    experiment_plan = example_data[index_ex]['experiment_plan']
+    return hypothesis, experiment_plan, hypothesis, experiment_plan
+
+##########  Phase 2 & 3  ##############
+def start_experiment_agent(hypothesis, plan):
+    if (not state_extract or not state_generate or not state_example):
+        return "", "", ""
+    global state_agent, step_index, state_complete
+    state_agent = True
+    step_index = 0
+    state_complete = False
+    # predefined_message = f"Implement the following hypothesis and experiment plan:\n\nHypothesis:\n{hypothesis}\n\nExperiment Plan:\n{plan}"
+    return example_data[index_ex]['code_init'], predefined_action_log, "", ""
+
+def submit_feedback(user_feedback, history, previous_response):
+    if (not state_extract or not  state_generate or not state_agent or not state_example):
+        return "", "", ""
+    global step_index, state_complete
+    step_index += 1
+    msg = history
+    if step_index < len(process_steps):
+        msg += previous_response + "\nUser feedback:" + user_feedback + "\n\n"
+        response_info = process_steps[step_index]
+        response = info_to_message(response_info)  # Convert dictionary to formatted string
+        response += "Please provide feedback based on the history, response entries, and observation, and questions: "
+        step_index += 1
+        msg += response
+    else:
+        state_complete = True
+        response = "Agent Finished."
+        
+    return msg, response, example_data[index_ex]['code_init'] if state_complete else example_data[index_ex]['code_final'], ""
+
+def load_phase_2_inputs(hypothesis, plan):
+    return hypothesis, plan, "# Code implementation will be displayed here after Start ExperimentAgent."
+
+
+
+predefined_action_log = """
+[Reasoning]: To understand the initial structure and functionality of train.py for effective improvements.
+[Action]: Inspect Script (train.py)
+Input: {"script_name": "train.py", "start_line_number": "1", "end_line_number": "74"}
+Objective: Understand the training script, including data processing, [...]
+[Observation]: The train.py script imports [...]. Sets random seeds [...]. Defines [...] Placeholder functions [...] exist without implementation. [...]
+[Feedback]: The script structure is clear, but key functions (train_model, predict) need proper implementation for proposed model training and prediction.\n
+"""
+
+
+predefined_observation = """
+Epoch [1/10],
+Train MSE: 0.543,
+Test MSE: 0.688
+Epoch [2/10],
+Train MSE: 0.242,
+Test MSE: 0.493\n
+"""
+
+# Initialize the global step_index and history
+process_steps = [
+    {
+        "Action": "Inspect Script Lines (train.py)",
+        "Observation": (
+            "The train.py script imports necessary libraries (e.g., pandas, sklearn, torch). "
+            "Sets random seeds for reproducibility. Defines compute_metrics_for_regression function "
+            "to calculate RMSE for different dimensions. Placeholder functions train_model and "
+            "predict exist without implementations."
+        ),
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The script executed successfully. Generated embeddings using the BERT model. Completed "
+            "the training process without errors. Metrics calculation placeholders indicated areas needing implementation."
+        ),
+    },
+    {
+        "Action": "Edit Script (train.py)",
+        "Observation": (
+            "Edited train.py to separate data loading, model definition, training loop, and evaluation into distinct functions. "
+            "The edited train.py now has clearly defined functions"
+            "for data loading (load_data), model definition (build_model), "
+            "training (train_model), and evaluation (evaluate_model). Similarly, eval.py is reorganized to load the model and perform predictions efficiently."
+        ),
+    },
+    {
+        "Action": "Retrieve Model",
+        "Observation": "CNN and BiLSTM retrieved.",
+    },
+    {
+        "Action": "Execute Script (train.py)",
+        "Observation": (
+            "The model trained over the specified number of epochs. Training and validation loss values are recorded for each epoch, "
+            "the decrease in loss indicates improved model performance."
+        )
+    },
+    {
+        "Action": "Evaluation",
+        "Observation": predefined_observation,
+    }
+]
+def info_to_message(info):
+    msg = ""
+    for k, v in info.items():
+        if isinstance(v, dict):
+            tempv = v
+            v = ""
+            for k2, v2 in tempv.items():
+                v += f"{k2}:\n  {v2}\n"
+        v = User.indent_text(v, 2)
+        msg += '-' * 64
+        msg += '\n'
+        msg += f"{k}:\n{v}\n"
+    return msg
+
+
+def handle_example_click(example_index):
+    global index_ex
+    index_ex = example_index
+    return load_example(index_ex)  # Simply return the text to display it in the textbox
+
+# Gradio Interface
+with gr.Blocks(css=".gr-examples-label {display: none;}", theme=gr.themes.Default()) as app:
+    gr.Markdown("# MLR- Copilot: Machine Learning Research based on LLM Agents")
+    gr.Markdown("###   ")
+    gr.Markdown("## <span style='color:Orange;'> This UI is for predefined example demo only.</span>")
+    gr.Markdown("## <span style='color:Orange;'> To reproduce the results please use [Github](https://github.com/du-nlp-lab/MLR-Copilot/)</span>")
+
+
+
+    gr.Markdown("MLR-Copilot is a framework where LLMs mimic researchers’ thought processes, designed to enhance the productivity of machine learning research by automating the generation and implementation of research ideas. It begins with a research paper, autonomously generating and validating these ideas, while incorporating human feedback to help reach executable research outcomes.")
+
+
+    gr.Markdown("##     ✅  Click an example at bottom to start  ⬇️")
+
+
+    # Use state variables to store generated hypothesis and experiment plan
+    hypothesis_state = gr.State("")
+    experiment_plan_state = gr.State("")
+
+    ########## Phase 1:  Research Idea Generation Tab ##############
+    with gr.Tab("💡Stage 1: Research Idea Generation"):
+        gr.Markdown("### Extract Research Elements and Generate Research Ideas")
+        with gr.Row():
+            with gr.Column():
+                paper_text_input = gr.Textbox(value="", lines=10, label="📑 Research Paper Text", interactive=False)
+
+                # extract_button = gr.Button("🔍 Extract Research Elements")
+                with gr.Row():
+                    tasks_output = gr.Textbox(placeholder="Research task definition", label="Research Tasks", lines=2, interactive=False)
+                    gaps_output = gr.Textbox(placeholder="Research gaps of current works", label="Research Gaps", lines=2, interactive=False)
+                    keywords_output = gr.Textbox(placeholder="Paper keywords", label="Keywords", lines=2, interactive=False)
+                    recent_works_output = gr.Textbox(placeholder="Recent works extracted from Semantic Scholar", label="Recent Works", lines=2, interactive=False)
+            with gr.Column():
+                with gr.Row():  # Move the button to the top 
+                    generate_button = gr.Button("✍️ Generate Research Hypothesis & Experiment Plan")
+                with gr.Group():
+                    gr.Markdown("### 🌟 Research Idea")
+                    with gr.Row():
+                        hypothesis_output = gr.Textbox(label="Generated Hypothesis", lines=20, interactive=False)
+                        experiment_plan_output = gr.Textbox(label="Generated Experiment Plan", lines=20, interactive=False)
+        
+    
+        # Step 1: Extract Research Elements
+        # extract_button.click(
+        #     fn=extract_research_elements, 
+        #     inputs=paper_text_input, 
+        #     outputs=[tasks_output, gaps_output, keywords_output, recent_works_output]
+        # )
+
+        generate_button.click(
+            fn=generate_and_store, 
+            inputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output], 
+            outputs=[hypothesis_output, experiment_plan_output, hypothesis_state, experiment_plan_state]
+        )
+
+        gr.Examples(
+            examples=example_text,
+            inputs=[paper_text_input],
+            outputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output],
+            fn=load_example_and_set_index,
+            run_on_click = True,
+            # label="⬇️ Click an example to load"
+        )
+        
+
+
+
+    ########## Phase 2 & 3:  Experiment implementation and execution ##############
+    with gr.Tab("🧪 Stage 2 & Stage 3: Experiment implementation and execution"):
+        gr.Markdown("### Interact with the ExperimentAgent")
+
+        with gr.Row():
+            with gr.Column():
+                with gr.Group():
+                    gr.Markdown("### 🌟 Generated Research Idea")
+                    with gr.Row():
+                        idea_input = gr.Textbox(label="Generated Research Hypothesis", lines=30, interactive=False)
+                        plan_input = gr.Textbox(label="Generated Experiment Plan", lines=30, interactive=False)
+        
+            with gr.Column():
+                start_exp_agnet = gr.Button("⚙️ Start / Reset ExperimentAgent", elem_classes=["agent-btn"])
+                with gr.Group():
+                    gr.Markdown("### Implementation + Execution Log")    
+                    log = gr.Textbox(label="📖 Execution Log", lines=20, interactive=False)
+                    code_display = gr.Code(label="🧑‍💻 Implementation", language="python", interactive=False)
+
+            with gr.Column():
+                response = gr.Textbox(label="🤖 ExperimentAgent Response", lines=30, interactive=False)
+                feedback = gr.Textbox(placeholder="N/A", label="🧑‍🔬 User Feedback", lines=3, interactive=True)
+                submit_button = gr.Button("Submit", elem_classes=["Submit-btn"])
+        
+        hypothesis_state.change(
+            fn=load_phase_2_inputs, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[idea_input, plan_input, code_display]
+        )
+        
+        # Start research agent
+        start_exp_agnet.click(
+            fn=start_experiment_agent, 
+            inputs=[hypothesis_state, experiment_plan_state], 
+            outputs=[code_display, log, response, feedback]
+        )
+        
+        submit_button.click(
+            fn=submit_feedback,
+            inputs=[feedback, log, response],
+            outputs=[log, response, code_display, feedback]
+        )
+
+# Test
+if __name__ == "__main__":
+    step_index = 0
+    app.launch(share=True)

app.py

@@ -24,6 +24,7 @@ example_text = [
     "Research Paper 2: An Empirical Study on the Impact of Code Review on Software Quality"
 ]
 
+
 # Load example JSON file
 def load_example_data():
     with open("example/example_data.json", "r") as json_file:
@@ -52,18 +53,23 @@ def load_example(example_id):
     index_ex = str(example_id)
     example = example_data[index_ex]
     paper_text = 'Title:\t' +  example['title'] + '\n\nAbstract:\t' + example['abstract']
-    return paper_text
+    tasks = example['research_tasks']
+    gaps = example['research_gaps']
+    keywords = example['keywords']
+    recent_works = "\n".join(example['recent_works'])
+    return [paper_text, tasks, gaps, keywords, recent_works]
+
+example_text = [load_example(1)[0], load_example(2)[0]]
 
-example_text = [load_example(1), load_example(2)]
 
 # Function to handle example clicks
 def load_example_and_set_index(paper_text_input):
     global index_ex, state_example
     state_example = True
     index_ex = str(example_text.index(paper_text_input) + 1)
-    paper_text =  load_example(index_ex)
+    example =  load_example(index_ex)
 
-    return paper_text,  "", "", "", "", "", ""
+    return example
 
 
 
@@ -88,7 +94,7 @@ def extract_research_elements(paper_text):
 
 # Step 2: Generate Research Hypothesis and Experiment Plan
 def generate_and_store(paper_text, tasks, gaps, keywords, recent_works):
-    if (not state_extract or not state_example or paper_text == ""):
+    if (not state_example):
         return "", "", "", ""
     global state_generate, index_ex
     state_generate = True
@@ -225,7 +231,7 @@ with gr.Blocks(css=".gr-examples-label {display: none;}", theme=gr.themes.Defaul
     gr.Markdown("MLR-Copilot is a framework where LLMs mimic researchers’ thought processes, designed to enhance the productivity of machine learning research by automating the generation and implementation of research ideas. It begins with a research paper, autonomously generating and validating these ideas, while incorporating human feedback to help reach executable research outcomes.")
 
 
-    gr.Markdown("## ⬇️Click an example at bottom to start")
+    gr.Markdown("##     ✅  Click an example at bottom to start  ⬇️")
 
 
     # Use state variables to store generated hypothesis and experiment plan
@@ -239,7 +245,7 @@ with gr.Blocks(css=".gr-examples-label {display: none;}", theme=gr.themes.Defaul
             with gr.Column():
                 paper_text_input = gr.Textbox(value="", lines=10, label="📑 Research Paper Text", interactive=False)
 
-                extract_button = gr.Button("🔍 Extract Research Elements")
+                # extract_button = gr.Button("🔍 Extract Research Elements")
                 with gr.Row():
                     tasks_output = gr.Textbox(placeholder="Research task definition", label="Research Tasks", lines=2, interactive=False)
                     gaps_output = gr.Textbox(placeholder="Research gaps of current works", label="Research Gaps", lines=2, interactive=False)
@@ -256,11 +262,11 @@ with gr.Blocks(css=".gr-examples-label {display: none;}", theme=gr.themes.Defaul
         
     
         # Step 1: Extract Research Elements
-        extract_button.click(
-            fn=extract_research_elements, 
-            inputs=paper_text_input, 
-            outputs=[tasks_output, gaps_output, keywords_output, recent_works_output]
-        )
+        # extract_button.click(
+        #     fn=extract_research_elements, 
+        #     inputs=paper_text_input, 
+        #     outputs=[tasks_output, gaps_output, keywords_output, recent_works_output]
+        # )
 
         generate_button.click(
             fn=generate_and_store, 
@@ -271,7 +277,7 @@ with gr.Blocks(css=".gr-examples-label {display: none;}", theme=gr.themes.Defaul
         gr.Examples(
             examples=example_text,
             inputs=[paper_text_input],
-            outputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output, hypothesis_output, experiment_plan_output],
+            outputs=[paper_text_input, tasks_output, gaps_output, keywords_output, recent_works_output],
             fn=load_example_and_set_index,
             run_on_click = True,
             # label="⬇️ Click an example to load"