Update app.py

app.py

@@ -1,63 +1,31 @@
 import gradio as gr
 from huggingface_hub import InferenceClient
-
-# Initialize Hugging Face Inference Client
-client = InferenceClient("HuggingFaceH4/zephyr-7b-beta")
-
-# Response Function
-def respond(message, history, system_message, max_tokens, temperature, top_p):
-    messages = [{"role": "system", "content": system_message}]
-
-    if isinstance(history, list):
-        for entry in history:
-            if isinstance(entry, dict):
-                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]})
-
-    messages.append({"role": "user", "content": message})
-
-    response = ""
-
-    for message in client.chat_completion(
-        messages,
-        max_tokens=max_tokens,
-        stream=True,
-        temperature=temperature,
-        top_p=top_p,
-    ):
-        token = message.choices[0].delta.content
-        response += token
-        yield response
-
-# Gradio Chat Interface
-demo = gr.ChatInterface(
-    respond,
-    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
 from transformers import AutoModelForCausalLM, AutoTokenizer, Trainer, TrainingArguments
-from datasets import Dataset
+from datasets import Dataset, load_dataset
 from peft import LoraConfig, get_peft_model
 import torch
 
+# Initialize Hugging Face Inference Client
+client = InferenceClient("HuggingFaceH4/zephyr-7b-beta")
+
 # Load GPT-2 model and tokenizer
 model_name = "gpt2"
 tokenizer = AutoTokenizer.from_pretrained(model_name)
 model = AutoModelForCausalLM.from_pretrained(model_name)
 
+# Add padding token (GPT-2 fix)
+if tokenizer.pad_token is None:
+    tokenizer.pad_token = tokenizer.eos_token
+
 # Custom Dataset (Predefined Q&A Pairs for Project Expo)
 custom_data = [
     {"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!"},
+    {"text": "Who invented the computer?", "label": "Charles Babbage is known as the father of the computer."},
+    {"text": "Tell me a joke.", "label": "Why don’t scientists trust atoms? Because they make up everything!"},
+    {"text": "Who is the Prime Minister of India?", "label": "The current Prime Minister of India is Narendra Modi."},
+    {"text": "Who created you?", "label": "I was created by an expert team specializing in AI fine-tuning and web development."}
 ]
 
 # Convert custom dataset to Hugging Face Dataset
@@ -66,20 +34,28 @@ dataset_custom = Dataset.from_dict({
     "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%]")
+
 # Tokenization function
 def tokenize_function(examples):
-    return tokenizer(examples["text"], truncation=True, padding="max_length", max_length=512)
+    return tokenizer(
+        examples["text"],
+        truncation=True,
+        padding="max_length",
+        max_length=512
+    )
 
 tokenized_datasets = dataset_custom.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"]
+    target_modules=["c_attn", "c_proj"]  # Apply LoRA to attention layers
 )
 
 model = get_peft_model(model, lora_config)
-model.gradient_checkpointing_enable()
+model.gradient_checkpointing_enable()  # Enable checkpointing for memory efficiency
 
 # Training arguments
 training_args = TrainingArguments(
@@ -115,6 +91,50 @@ 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()
+# Gradio Chat Interface
+def respond(message, history, system_message, max_tokens, temperature, top_p):
+    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
+            elif isinstance(entry, tuple) and len(entry) == 2:
+                messages.append({"role": "user", "content": entry[0]})
+                messages.append({"role": "assistant", "content": entry[1]})
+
+    messages.append({"role": "user", "content": message})
+    response = ""
+
+    for message in client.chat_completion(
+        messages,
+        max_tokens=max_tokens,
+        stream=True,
+        temperature=temperature,
+        top_p=top_p,
+    ):
+        token = message.choices[0].delta.content
+        response += token
+        yield response
+
+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 (nucleus sampling)",
+        ),
+    ],
+)
+
+# Launch the Gradio app
+if __name__ == "__main__":
+    demo.launch()