Update app.py

app.py

@@ -4,28 +4,20 @@ from huggingface_hub import InferenceClient
 # Initialize Hugging Face Inference Client
 client = InferenceClient("HuggingFaceH4/zephyr-7b-beta")
 
-# Response Function (Now Compatible with 'type=messages')
-def respond(
-    message,
-    history,
-    system_message,
-    max_tokens,
-    temperature,
-    top_p,
-):
-    # Correct structure for Gradio's 'messages' format
+# Response Function
+def respond(message, history, system_message, max_tokens, temperature, top_p):
+    # Ensure correct message structure
     messages = [{"role": "system", "content": system_message}]
 
-    # Ensure 'history' is handled as a list of dicts
     if isinstance(history, list):
         for entry in history:
             if isinstance(entry, dict):
-                messages.append(entry)  # Correct format already
+                messages.append(entry)
             elif isinstance(entry, tuple) and len(entry) == 2:
                 messages.append({"role": "user", "content": entry[0]})
                 messages.append({"role": "assistant", "content": entry[1]})
 
-    # Add current user message
+    # Append user message
     messages.append({"role": "user", "content": message})
 
     # Initialize response
@@ -44,13 +36,21 @@ def respond(
         yield response
 
 
-# Launch the Gradio app
-if __name__ == "__main__":
-    demo.launch()
+# Gradio Chat Interface
+demo = gr.ChatInterface(
+    respond,
+    chatbot=gr.Chatbot(type="messages"),
+    additional_inputs=[
+        gr.Textbox(value="You are a friendly Chatbot.", label="System message"),
+        gr.Slider(minimum=1, maximum=2048, value=512, step=1, label="Max new tokens"),
+        gr.Slider(minimum=0.1, maximum=4.0, value=0.7, step=0.1, label="Temperature"),
+        gr.Slider(minimum=0.1, maximum=1.0, value=0.95, step=0.05, label="Top-p"),
+    ],
+)
 
-# Fine-Tuning GPT-2 on Hugging Face Spaces (Streaming 40GB Dataset, No Storage Issues)
+# Fine-Tuning GPT-2 on Hugging Face Spaces
 from transformers import AutoModelForCausalLM, AutoTokenizer, Trainer, TrainingArguments
-from datasets import load_dataset
+from datasets import Dataset
 from peft import LoraConfig, get_peft_model
 import torch
 
@@ -65,20 +65,17 @@ model = AutoModelForCausalLM.from_pretrained(model_name)
 
 # Custom Dataset (Predefined Q&A Pairs for Project Expo)
 custom_data = [
-    {"prompt": "Who are you?", "response": "I am Eva, a virtual voice assistant."},
-    {"prompt": "What is your name?", "response": "I am Eva, how can I help you?"},
-    {"prompt": "What can you do?", "response": "I can assist with answering questions, searching the web, and much more!"},
-    {"prompt": "Who invented the computer?", "response": "Charles Babbage is known as the father of the computer."},
-    {"prompt": "Tell me a joke.", "response": "Why don’t scientists trust atoms? Because they make up everything!"},
-    {"prompt": "Who is the Prime Minister of India?", "response": "The current Prime Minister of India is Narendra Modi."},
-    {"prompt": "Who created you?", "response": "I was created by an expert team specializing in AI fine-tuning and web development."}
+    {"text": "Who are you?", "label": "I am Eva, a virtual voice assistant."},
+    {"text": "What is your name?", "label": "I am Eva, how can I help you?"},
+    {"text": "What can you do?", "label": "I can assist with answering questions, searching the web, and much more!"},
 ]
 
 # Convert custom dataset to Hugging Face Dataset
-dataset_custom = load_dataset("json", data_files={"train": custom_data})
+dataset_custom = Dataset.from_dict({"text": [d['text'] for d in custom_data], 
+                                    "label": [d['label'] for d in custom_data]})
 
-# Load OpenWebText dataset (5% portion to avoid streaming issues)
-dataset = load_dataset("Skylion007/openwebtext", split="train[:20%]")
+# Load OpenWebText dataset (5% portion)
+dataset = dataset_custom.train_test_split(test_size=0.2)['train']
 
 # Tokenization function
 def tokenize_function(examples):
@@ -89,11 +86,11 @@ tokenized_datasets = dataset.map(tokenize_function, batched=True)
 # Apply LoRA for efficient fine-tuning
 lora_config = LoraConfig(
     r=8, lora_alpha=32, lora_dropout=0.05, bias="none",
-    target_modules=["c_attn", "c_proj"]  # Apply LoRA to attention layers
+    target_modules=["c_attn", "c_proj"]
 )
 
 model = get_peft_model(model, lora_config)
-model.gradient_checkpointing_enable()  # Enable checkpointing for memory efficiency
+model.gradient_checkpointing_enable()
 
 # Training arguments
 training_args = TrainingArguments(
@@ -129,5 +126,6 @@ def generate_response(prompt):
     outputs = model.generate(**inputs, max_length=100)
     return tokenizer.decode(outputs[0], skip_special_tokens=True)
 
+# Corrected Gradio Interface
 demo = gr.Interface(fn=generate_response, inputs="text", outputs="text")
 demo.launch()