app.py

import gradio as gr
from llama_cpp import Llama
from huggingface_hub import hf_hub_download

# Function to load a user-specified model from Hugging Face
def load_user_model(repo_id, model_file):
    print(f"Downloading model {model_file} from repository {repo_id}...")
    local_path = hf_hub_download(repo_id=repo_id, filename=model_file)
    print(f"Model downloaded to: {local_path}")
    return Llama(model_path=local_path, n_ctx=2048, n_threads=8)

# Generate a response using the specified model and prompt
def generate_response(model, prompt):
    response = model(prompt, max_tokens=256, temperature=0.7)
    return response["choices"][0]["text"]

# Evaluate responses generated by two models using the LoRA evaluation model
def evaluate_responses(prompt, repo_a, model_a, repo_b, model_b, criteria_list):
    # Load user-specified models
    model_a_instance = load_user_model(repo_a, model_a)
    model_b_instance = load_user_model(repo_b, model_b)
    
    # Generate responses
    response_a = generate_response(model_a_instance, prompt)
    response_b = generate_response(model_b_instance, prompt)
    
    print(f"Response A: {response_a}")
    print(f"Response B: {response_b}")
    
    # Format the evaluation prompt for the LoRA model
    evaluation_prompt = f"""
Prompt: {prompt}

Response A: {response_a}
Response B: {response_b}

Evaluation Criteria: {criteria_list}

Please evaluate the responses based on the selected criteria. For each criterion, rate both responses on a scale from 1 to 10 and provide a justification. Finally, declare the winner (or 'draw' if they are equal).
"""
    # Use the LoRA model to evaluate the responses
    evaluation_response = lora_model.create_completion(
        prompt=evaluation_prompt,
        max_tokens=512,
        temperature=0.5
    )
    evaluation_results = evaluation_response["choices"][0]["text"]

    final_output = f"""
    Response A:\n{response_a}\n\n
    Response B:\n{response_b}\n\n
    Evaluation Results:\n{evaluation_results}
    """
    return final_output

# Load the base LoRA evaluation model
def load_lora_model():
    repo_id = "KolumbusLindh/LoRA-4100"
    model_file = "unsloth.F16.gguf"
    print(f"Downloading LoRA evaluation model from repository {repo_id}...")
    local_path = hf_hub_download(repo_id=repo_id, filename=model_file)
    print(f"LoRA evaluation model downloaded to: {local_path}")
    return Llama(model_path=local_path, n_ctx=2048, n_threads=8)

lora_model = load_lora_model()
print("LoRA evaluation model loaded successfully!")

# Gradio interface
with gr.Blocks(title="LLM as a Judge") as demo:
    gr.Markdown("## LLM as a Judge 🧐")

    # Inputs for Model A repository and file
    repo_a_input = gr.Textbox(label="Model A Repository (e.g., KolumbusLindh/LoRA-4100)", placeholder="Enter the Hugging Face repo name for Model A...")
    model_a_input = gr.Textbox(label="Model A File Name (e.g., unsloth.F16.gguf)", placeholder="Enter the model filename for Model A...")

    # Inputs for Model B repository and file
    repo_b_input = gr.Textbox(label="Model B Repository (e.g., KolumbusLindh/LoRA-4100)", placeholder="Enter the Hugging Face repo name for Model B...")
    model_b_input = gr.Textbox(label="Model B File Name (e.g., unsloth.F16.gguf)", placeholder="Enter the model filename for Model B...")

    # Input for prompt and evaluation criteria
    prompt_input = gr.Textbox(label="Enter Prompt", placeholder="Enter the prompt here...", lines=3)
    criteria_dropdown = gr.CheckboxGroup(
        label="Select Up to 3 Evaluation Criteria",
        choices=["Clarity", "Completeness", "Accuracy", "Relevance", "User-Friendliness", "Depth", "Creativity"],
        value=["Clarity"],
        max_choices=3
    )

    # Button to evaluate responses
    evaluate_button = gr.Button("Evaluate Models")
    
    # Output for evaluation results
    evaluation_output = gr.Textbox(
        label="Evaluation Results",
        placeholder="The evaluation results will appear here...",
        lines=10,
        interactive=False
    )

    # Link the evaluation function to the button
    evaluate_button.click(
        fn=evaluate_responses,
        inputs=[prompt_input, repo_a_input, model_a_input, repo_b_input, model_b_input, criteria_dropdown],
        outputs=[evaluation_output]
    )

# Launch the Gradio app
if __name__ == "__main__":
    demo.launch()  # Add share=True to create a public link