app.py

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

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

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

# Function to load a user-specified model
def load_user_model(model_path):
    print(f"Loading user model from: {model_path}")
    return Llama(model_path=model_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 model
def evaluate_responses(prompt, model_a_path, model_b_path, evaluation_criteria):
    # Load user-specified models
    model_a = load_user_model(model_a_path)
    model_b = load_user_model(model_b_path)
    
    # Generate responses
    response_a = generate_response(model_a, prompt)
    response_b = generate_response(model_b, 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: {evaluation_criteria}

Please evaluate the responses based on the criteria above. Rate each response on a scale from 1 to 10 for each criterion and provide a detailed explanation. Finally, declare a winner or state '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
    )
    return evaluation_response["choices"][0]["text"]

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

    # User inputs for models, prompt, and evaluation criteria
    model_a_input = gr.Textbox(label="Model A Path or URL", placeholder="Enter the path or URL for Model A...")
    model_b_input = gr.Textbox(label="Model B Path or URL", placeholder="Enter the path or URL for Model B...")
    prompt_input = gr.Textbox(label="Enter Prompt", placeholder="Enter the prompt here...", lines=3)
    criteria_dropdown = gr.Dropdown(
        label="Select Evaluation Criteria",
        choices=["Clarity", "Completeness", "Accuracy", "Relevance", "User-Friendliness", "Depth", "Creativity"],
        value="Clarity",
        type="value"
    )
    evaluate_button = gr.Button("Evaluate Models")
    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, model_a_input, model_b_input, criteria_dropdown],
        outputs=[evaluation_output]
    )

# Launch the Gradio app
if __name__ == "__main__":
    demo.launch()