import gradio as gr import requests import os import random import uuid from datetime import datetime import csv from huggingface_hub import HfApi # This is a placeholder for the Hugging Face token. # In a real environment, you would get this from a secure location. HF_TOKEN = os.getenv("HF_Key") # Three models: base, math, general (from environment variables set in HF Space secrets) MODEL_INFOS = [ { "name": "Base", "endpoint_url": os.getenv("BASE_ENDPOINT_URL", "https://o3pz2i9x2k6otr2a.eu-west-1.aws.endpoints.huggingface.cloud/v1/"), "model": os.getenv("BASE_ENDPOINT_MODEL", "qwen3-4b-instruct-2507-pxe") }, { "name": "Math", "endpoint_url": os.getenv("MATH_ENDPOINT_URL", "https://jockj5ko30gpg5lg.eu-west-1.aws.endpoints.huggingface.cloud/v1/"), "model": os.getenv("MATH_ENDPOINT_MODEL", "teach-math-qwen3-4b-2507-r1--uab") }, { "name": "General", "endpoint_url": os.getenv("GENERAL_ENDPOINT_URL", "https://ucewop5x3jsguqwq.eu-west-1.aws.endpoints.huggingface.cloud/v1/"), "model": os.getenv("GENERAL_ENDPOINT_MODEL", "CanisAI/teach-generalist-qwen3-4b-2507-r1-merged") }, ] # Dataset repo ID - change this to your actual dataset repo DATASET_REPO_ID = "CanisAI/mvlg-data" api = HfApi() # Feedback points (sliders 1-10) - education-specific criteria FEEDBACK_POINTS = [ "How clear was the explanation?", "How helpful were the steps in guiding you to the solution?", "How well did the assistant adapt to your learning style?", "How motivating and encouraging was the response?", "How accurate and reliable was the information provided?", "How relevant was the information to your question?", "How natural and conversational was the interaction?", "How much do you trust the assistant?" ] def query_chat_endpoint(endpoint_url, model, messages, max_tokens=150, temperature=0.7): url = endpoint_url.rstrip("/") + "/chat/completions" headers = { "Accept": "application/json", "Content-Type": "application/json", "Authorization": f"Bearer {HF_TOKEN}" } payload = { "model": model, "messages": messages, "max_tokens": max_tokens, "temperature": temperature, "stream": False } try: response = requests.post(url, headers=headers, json=payload) response.raise_for_status() result = response.json() return result["choices"][0]["message"]["content"] except Exception as e: return f"Error: {str(e)}" def chat_multi_llm(message, history, current_model_state, conversation_id_state): if history is None: history = [] if current_model_state is None: current_model_state = random.choice(MODEL_INFOS) conversation_id_state = str(uuid.uuid4()) messages = [{"role": msg["role"], "content": msg["content"]} for msg in history] messages.append({"role": "user", "content": message}) model_name = current_model_state["name"] endpoint_url = current_model_state["endpoint_url"] model = current_model_state["model"] answer = query_chat_endpoint(endpoint_url, model, messages) log_chat_to_csv(message, history, {model_name: answer}, model_name, conversation_id_state) new_history = history + [{"role": "user", "content": message}, {"role": "assistant", "content": answer}] return new_history, current_model_state, conversation_id_state def log_chat_to_csv(message, history, results, used_model, conversation_id, filename="chat_conversations_log.csv"): from os.path import isfile file_exists = isfile(filename) with open(filename, mode="a", encoding="utf-8", newline="") as csvfile: writer = csv.writer(csvfile) if not file_exists: header = ["timestamp", "conversation_id", "history", "user_message", "used_model", "response"] writer.writerow(header) row = [datetime.now().isoformat(), conversation_id, str(history), message, used_model, list(results.values())[0]] writer.writerow(row) # Upload to Hugging Face dataset repo try: api.upload_file( path_or_fileobj=filename, path_in_repo=filename, repo_id=DATASET_REPO_ID, repo_type="dataset", token=HF_TOKEN ) except Exception as e: print(f"Error uploading to HF dataset: {e}") def submit_feedback(current_model, conversation_id, *slider_values): filename = "feedback_log.csv" from os.path import isfile file_exists = isfile(filename) with open(filename, "a", encoding="utf-8", newline="") as f: writer = csv.writer(f) if not file_exists: writer.writerow(["timestamp", "conversation_id", "used_model"] + FEEDBACK_POINTS) writer.writerow([ datetime.now().isoformat(), conversation_id, current_model["name"] if current_model else "None" ] + list(slider_values)) # Upload feedback to HF dataset repo try: api.upload_file( path_or_fileobj=filename, path_in_repo=filename, repo_id=DATASET_REPO_ID, repo_type="dataset", token=HF_TOKEN ) except Exception as e: print(f"Error uploading feedback to HF dataset: {e}") return ( gr.update(visible=False), # feedback_col gr.update(visible=True), # chat_col gr.update(value="Thank you! You can start a new conversation.", visible=True), # feedback_info [], # history_state - reset None, # current_model_state - reset None, # conversation_id_state - reset gr.update(interactive=True), # msg gr.update(interactive=True), # submit_btn [], # chatbot - reset gr.update(visible=False) # end_info - hide ) with gr.Blocks() as demo: gr.Markdown(""" # LLM Case Study: Multi-Model Chat Comparison Start a conversation. After finishing, you can provide feedback and start a new conversation. By Using the app you accept that your interactions and feedback will be logged and used for research purposes. Please don't share any personal, sensitive, or confidential information. """) history_state = gr.BrowserState([]) # persists across page refreshes # Persist the selected model and conversation id in the browser so they # survive page refreshes. Using BrowserState prevents the model from # being re-selected randomly mid-conversation on reload. current_model_state = gr.BrowserState(None) conversation_id_state = gr.BrowserState(None) with gr.Column(visible=True) as chat_col: chatbot = gr.Chatbot(type="messages", value=[]) msg = gr.Textbox(placeholder="Enter your message...", show_label=False) submit_btn = gr.Button("Send") end_btn = gr.Button("End conversation and give feedback") end_info = gr.Markdown("", visible=False) with gr.Column(visible=False) as feedback_col: sliders = [gr.Slider(1, 10, value=5, step=1, label=label) for label in FEEDBACK_POINTS] feedback_btn = gr.Button("Submit feedback and start new conversation") feedback_info = gr.Markdown("", visible=False) def user_message(message, history, current_model, conversation_id): if message is None or message.strip() == "": return history, "", history, current_model, conversation_id new_history, updated_model, updated_conv_id = chat_multi_llm(message, history, current_model, conversation_id) return new_history, "", new_history, updated_model, updated_conv_id def load_chat_history(history): """Load the chat history into the chatbot on page load""" if history is None: return [] return history def end_conversation(): return ( gr.update(visible=True), # feedback_col gr.update(visible=False), # chat_col gr.update(value="Please provide feedback on the last conversation.", visible=True), # end_info gr.update(interactive=False), # msg gr.update(interactive=False) # submit_btn ) msg.submit( user_message, inputs=[msg, history_state, current_model_state, conversation_id_state], outputs=[chatbot, msg, history_state, current_model_state, conversation_id_state], queue=False ) submit_btn.click( user_message, inputs=[msg, history_state, current_model_state, conversation_id_state], outputs=[chatbot, msg, history_state, current_model_state, conversation_id_state], queue=False ) end_btn.click( end_conversation, inputs=None, outputs=[feedback_col, chat_col, end_info, msg, submit_btn] ) feedback_btn.click( submit_feedback, inputs=[current_model_state, conversation_id_state] + sliders, outputs=[feedback_col, chat_col, feedback_info, history_state, current_model_state, conversation_id_state, msg, submit_btn, chatbot, end_info] ) # Load chat history from BrowserState on page load demo.load( load_chat_history, inputs=[history_state], outputs=[chatbot] ) if __name__ == "__main__": demo.launch()