import gradio as gr import torch from transformers import AutoTokenizer, AutoModelForCausalLM import evaluate import re import matplotlib matplotlib.use('Agg') import matplotlib.pyplot as plt import io import base64 import os from huggingface_hub import login import spaces # Read token and login hf_token = os.getenv("HF_TOKEN_READ_WRITE") if hf_token: login(hf_token) else: print("⚠️ No HF_TOKEN_READ_WRITE found in environment") # --------------------------------------------------------------------------- # 1. Model and tokenizer setup # --------------------------------------------------------------------------- model_name = "mistralai/Mistral-7B-Instruct-v0.3" tokenizer = None model = None @spaces.GPU def load_model(): global tokenizer, model if tokenizer is None: tokenizer = AutoTokenizer.from_pretrained(model_name, token=hf_token) if model is None: model = AutoModelForCausalLM.from_pretrained( model_name, token=hf_token, torch_dtype=torch.float16 ) model.to('cuda') return model, tokenizer # --------------------------------------------------------------------------- # 2. Test dataset # --------------------------------------------------------------------------- test_data = [ {"question": "What is 2+2?", "answer": "4"}, {"question": "What is 3*3?", "answer": "9"}, {"question": "What is 10/2?", "answer": "5"}, ] # --------------------------------------------------------------------------- # 3. Load metric # --------------------------------------------------------------------------- accuracy_metric = evaluate.load("accuracy") # --------------------------------------------------------------------------- # 4. Inference helper functions # --------------------------------------------------------------------------- @spaces.GPU def generate_answer(question): """ Generates an answer using Mistral's instruction format. """ model, tokenizer = load_model() # Mistral instruction format prompt = f"""[INST] {question}. Provide only the numerical answer. [/INST]""" inputs = tokenizer(prompt, return_tensors="pt").to('cuda') with torch.no_grad(): outputs = model.generate( **inputs, max_new_tokens=50, temperature=0.0, # deterministic pad_token_id=tokenizer.pad_token_id, eos_token_id=tokenizer.eos_token_id ) text_output = tokenizer.decode(outputs[0], skip_special_tokens=True) # Remove the original question from the output return text_output.replace(question, "").strip() def parse_answer(model_output): """ Extract numeric answer from model's text output. """ # Look for numbers (including decimals) match = re.search(r"(-?\d*\.?\d+)", model_output) if match: return match.group(1) return model_output.strip() # --------------------------------------------------------------------------- # 5. Evaluation routine # --------------------------------------------------------------------------- @spaces.GPU(duration=120) # Allow up to 2 minutes for full evaluation def run_evaluation(): predictions = [] references = [] raw_outputs = [] # Store full model outputs for display for sample in test_data: question = sample["question"] reference_answer = sample["answer"] # Model inference model_output = generate_answer(question) predicted_answer = parse_answer(model_output) predictions.append(predicted_answer) references.append(reference_answer) raw_outputs.append({ "question": question, "model_output": model_output, "parsed_answer": predicted_answer, "reference": reference_answer }) # Normalize answers def normalize_answer(ans): return str(ans).lower().strip() norm_preds = [normalize_answer(p) for p in predictions] norm_refs = [normalize_answer(r) for r in references] # Compute accuracy results = accuracy_metric.compute(predictions=norm_preds, references=norm_refs) accuracy = results["accuracy"] # Create visualization fig, ax = plt.subplots(figsize=(8, 6)) correct_count = sum(p == r for p, r in zip(norm_preds, norm_refs)) incorrect_count = len(test_data) - correct_count bars = ax.bar(["Correct", "Incorrect"], [correct_count, incorrect_count], color=["#2ecc71", "#e74c3c"]) # Add value labels on bars for bar in bars: height = bar.get_height() ax.text(bar.get_x() + bar.get_width()/2., height, f'{int(height)}', ha='center', va='bottom') ax.set_title("Evaluation Results") ax.set_ylabel("Count") ax.set_ylim([0, len(test_data) + 0.5]) # Convert plot to base64 buf = io.BytesIO() plt.savefig(buf, format="png", bbox_inches='tight', dpi=300) buf.seek(0) plt.close(fig) data = base64.b64encode(buf.read()).decode("utf-8") # Create detailed results HTML details_html = """

Detailed Results:

""" for result in raw_outputs: details_html += f""" """ details_html += "
Question Model Output Parsed Answer Reference
{result['question']} {result['model_output']} {result['parsed_answer']} {result['reference']}
" full_html = f"""
{details_html}
""" return f"Accuracy: {accuracy:.2f}", full_html # --------------------------------------------------------------------------- # 6. Gradio Interface # --------------------------------------------------------------------------- with gr.Blocks() as demo: gr.Markdown("# Mistral-7B Math Evaluation Demo") gr.Markdown(""" This demo evaluates Mistral-7B on basic math problems. Press the button below to run the evaluation. """) eval_button = gr.Button("Run Evaluation", variant="primary") output_text = gr.Textbox(label="Results") output_plot = gr.HTML(label="Visualization and Details") eval_button.click( fn=run_evaluation, inputs=None, outputs=[output_text, output_plot] ) demo.launch()