app.py

# app.py
import gradio as gr
import pandas as pd
from huggingface_hub import InferenceClient

# Initialize the free Hugging Face Inference Client
# When deployed on HF Spaces, this automatically leverages the environment to run key-free!
client = InferenceClient()

def run_evaluation(csv_file, eval_type):
    if csv_file is None:
        return "Please upload a CSV file first.", None
    
    # 1. Load the dataset (Expects 'prompt' and 'response' columns)
    df = pd.read_csv(csv_file.name)
    if not all(col in df.columns for col in ['prompt', 'response']):
        return "Error: CSV must contain at least 'prompt' and 'response' columns.", None

    results = []
    
    # 2. Run the evaluations
    for idx, row in df.iterrows():
        prompt = row['prompt']
        response = row['response']
        score = 0
        reasoning = ""

        if eval_type == "Basic: Length & Overlap":
            # Pure local python check: check word count and keyword overlap
            word_count = len(response.split())
            overlap_words = set(prompt.lower().split()) & set(response.lower().split())
            score = round(min(len(overlap_words) / max(len(prompt.split()), 1) * 5, 5), 1)
            reasoning = f"Length: {word_count} words. Prompt keyword overlap: {len(overlap_words)} words."
            
        elif eval_type == "LLM-as-a-Judge: Correctness (Free Llama-3)":
            # Use a free open-source model as our evaluator
            judge_prompt = f"""<|im_start|>system
You are a precise evaluation assistant. Score the response based on the prompt.
Output your final answer exactly in this format:
Score: [Your Score from 1 to 5]
Reasoning: [One brief sentence explaining the score]
<|im_end|>
<|im_start|>user
Prompt: {prompt}
Response: {response}
<|im_end|>
<|im_start|>assistant
"""
            try:
                # We use Qwen-2.5-72B-Instruct or Llama-3-8B-Instruct (both are highly capable and free)
                completion = client.text_generation(
                    prompt=judge_prompt,
                    model="Qwen/Qwen2.5-72B-Instruct",
                    max_new_tokens=150,
                    temperature=0.1
                )
                
                # Parse the judge's response
                for line in completion.split('\n'):
                    if "Score:" in line:
                        # Extract number
                        score_part = line.split("Score:")[-1].strip()
                        # Clean up any trailing text
                        score = float(''.join(c for c in score_part if c.isdigit() or c == '.'))
                    elif "Reasoning:" in line:
                        reasoning = line.split("Reasoning:")[-1].strip()
                        
            except Exception as e:
                score, reasoning = 0, f"HF API Error: {str(e)}"

        results.append({"Score": score, "Reasoning": reasoning})

    # 3. Compile and present results
    results_df = pd.concat([df, pd.DataFrame(results)], axis=1)
    avg_score = results_df["Score"].mean()
    summary_text = f"### Evaluation Complete! \n**Average Score:** {avg_score:.2f} / 5.0"
    
    return summary_text, results_df

# --- Gradio UI Layout ---
with gr.Blocks(title="Free LLM Eval MVP") as demo:
    gr.Markdown("# 🧪 Keyless LLM Eval Engine")
    gr.Markdown("Upload a CSV dataset with your model's outputs and evaluate them using free open-source models.")
    
    with gr.Row():
        with gr.Column(scale=1):
            file_input = gr.File(label="Upload Evaluation CSV (must have 'prompt' and 'response')", file_types=[".csv"])
            eval_selector = gr.Radio(
                choices=["Basic: Length & Overlap", "LLM-as-a-Judge: Correctness (Free Llama-3)"],
                value="Basic: Length & Overlap",
                label="Select Evaluator"
            )
            submit_btn = gr.Button("Run Evals", variant="primary")
            
        with gr.Column(scale=2):
            output_summary = gr.Markdown()
            output_table = gr.DataFrame(label="Evaluation Results")

    submit_btn.click(
        fn=run_evaluation, 
        inputs=[file_input, eval_selector], 
        outputs=[output_summary, output_table]
    )

if __name__ == "__main__":
    demo.launch()