Leetmonkey In Action. Darn LeetMonkey these days

app.py

@@ -0,0 +1,117 @@
+import gradio as gr
+from llama_cpp import Llama
+import re
+from datasets import load_dataset
+from transformers import AutoTokenizer
+import autopep8
+import random
+import textwrap
+
+# Load the dataset
+dataset = load_dataset("sugiv/leetmonkey_python_dataset")
+val_dataset = dataset["train"].train_test_split(test_size=0.1)["test"]
+
+# Load the tokenizer
+tokenizer = AutoTokenizer.from_pretrained("sugiv/Leetmonkey-peft")
+
+# GGUF model options
+gguf_models = {
+    "Exact Copy": "sugiv/leetmonkey-peft-gguf/leetmonkey_peft_exact_copy.gguf",
+    "F16": "sugiv/leetmonkey-peft-gguf/leetmonkey_peft_f16.gguf",
+    "Q8_0": "sugiv/leetmonkey-peft-gguf/leetmonkey_peft__q8_0.gguf",
+    "Super Block Q6": "sugiv/leetmonkey-peft-gguf/leetmonkey_peft_super_block_q6.gguf"
+}
+
+generation_kwargs = {
+    "max_tokens": 2048,
+    "stop": ["```", "### Instruction:", "### Response:"],
+    "echo": False,
+    "temperature": 0.2,
+    "top_k": 50,
+    "top_p": 0.95,
+    "repeat_penalty": 1.1
+}
+
+def run_llama_cpp(instruction, model_path):
+    llm = Llama(
+        model_path=model_path,
+        n_ctx=2048,
+        n_threads=4,
+        n_gpu_layers=0,
+        verbose=False
+    )
+    
+    system_prompt = "You are a Python coding assistant specialized in solving LeetCode problems. Provide only the complete implementation of the given function. Ensure proper indentation and formatting. Do not include any explanations or multiple solutions."
+    full_prompt = f"""<｜begin▁of▁sentence｜>
+### Instruction:
+{system_prompt}
+
+Implement the following function for the LeetCode problem:
+
+{instruction}
+
+### Response:
+Here's the complete Python function implementation:
+
+```python
+"""
+    
+    res = llm(full_prompt, **generation_kwargs)
+    return res["choices"][0]["text"]
+
+def extract_and_format_code(text):
+    code_match = re.search(r'```python\s*(.*?)\s*```', text, re.DOTALL)
+    if code_match:
+        code = code_match.group(1)
+    else:
+        code = text
+
+    code = re.sub(r'^.*?(?=def\s+\w+\s*\()', '', code, flags=re.DOTALL)
+    code = textwrap.dedent(code)
+    lines = code.split('\n')
+    func_def_index = next((i for i, line in enumerate(lines) if line.strip().startswith('def ')), 0)
+    indented_lines = [lines[func_def_index]]
+    for line in lines[func_def_index + 1:]:
+        if line.strip():
+            indented_lines.append('    ' + line)
+        else:
+            indented_lines.append(line)
+
+    formatted_code = '\n'.join(indented_lines)
+
+    try:
+        return autopep8.fix_code(formatted_code)
+    except:
+        return formatted_code
+
+def select_random_problem():
+    sample = random.choice(val_dataset)
+    return sample['instruction']
+
+def generate_solution(problem, model_name):
+    model_path = gguf_models[model_name]
+    generated_output = run_llama_cpp(problem, model_path)
+    formatted_code = extract_and_format_code(generated_output)
+    return formatted_code
+
+def gradio_interface(problem, model_name):
+    solution = generate_solution(problem, model_name)
+    return solution
+
+with gr.Blocks() as demo:
+    gr.Markdown("# LeetCode Problem Solver")
+    
+    with gr.Row():
+        with gr.Column():
+            problem_display = gr.Textbox(label="LeetCode Problem", lines=10)
+            select_problem_btn = gr.Button("Select Random Problem")
+        
+        with gr.Column():
+            model_dropdown = gr.Dropdown(choices=list(gguf_models.keys()), label="Select GGUF Model", value="Exact Copy")
+            solution_display = gr.Code(label="Generated Solution", language="python")
+            generate_btn = gr.Button("Generate Solution")
+    
+    select_problem_btn.click(select_random_problem, outputs=problem_display)
+    generate_btn.click(gradio_interface, inputs=[problem_display, model_dropdown], outputs=solution_display)
+
+demo.launch()

requirements.txt

@@ -0,0 +1,5 @@
+gradio
+llama-cpp-python
+datasets
+transformers
+autopep8