app.py

import gradio as gr
from huggingface_hub import list_models
from sentence_transformers import SentenceTransformer, util
import numpy as np

# Load sentence transformer model for similarity calculation
semantic_model = SentenceTransformer('all-MiniLM-L6-v2')

# Function to fetch models from Hugging Face based on dynamic task filter
def fetch_models_from_hf(task_filter, limit=10):
    models = list_models(filter=task_filter, limit=limit)
    model_data = [
        {
            "model_id": model.modelId,
            "tags": model.tags,
            "downloads": model.downloads,
            "likes": model.likes,
            "last_modified": model.lastModified  # You could use this for recency
        }
        for model in models
    ]
    return model_data

# Normalize values for a 0-1 range
def normalize(values):
    min_val, max_val = min(values), max(values)
    return [(v - min_val) / (max_val - min_val) if max_val > min_val else 0 for v in values]

# Get weighted recommendations based on user query and additional metrics
def get_weighted_recommendations_from_hf(user_query, task_filter, weights=None):
    if weights is None:
        weights = {"similarity": 0.7, "downloads": 0.2, "likes": 0.1}  # Adjustable

    model_data = fetch_models_from_hf(task_filter)
    
    model_ids = [model["model_id"] for model in model_data]
    model_tags = [' '.join(model["tags"]) for model in model_data]
    
    model_embeddings = semantic_model.encode(model_tags)
    user_embedding = semantic_model.encode(user_query)
    
    similarities = util.pytorch_cos_sim(user_embedding, model_embeddings)[0].numpy()
    
    downloads = normalize([model["downloads"] for model in model_data])
    likes = normalize([model["likes"] for model in model_data])
    
    final_scores = []
    for i in range(len(model_data)):
        score = (
            weights["similarity"] * similarities[i] +
            weights["downloads"] * downloads[i] +
            weights["likes"] * likes[i]
        )
        final_scores.append((model_ids[i], score, similarities[i], downloads[i], likes[i]))
    
    ranked_recommendations = sorted(final_scores, key=lambda x: x[1], reverse=True)
    
    result = []
    for rank, (model_id, final_score, sim, downloads, likes) in enumerate(ranked_recommendations, 1):
        result.append(f"Rank {rank}: Model ID: {model_id}, Final Score: {final_score:.4f}, "
                      f"Similarity: {sim:.4f}, Downloads: {downloads:.4f}, Likes: {likes:.4f}")
    
    return '\n'.join(result)

# Define a Gradio interface function
def chatbot_interface(user_query, task_filter):
    return get_weighted_recommendations_from_hf(user_query, task_filter)

# Gradio Interface
interface = gr.Interface(
    fn=chatbot_interface,
    inputs=[
        gr.inputs.Textbox(label="Enter your query", placeholder="What kind of model or tag are you looking for?"),
        gr.inputs.Textbox(label="Task Filter (e.g., text-classification, summarization, atari)", placeholder="Enter the task"),
    ],
    outputs="text",
    title="Hugging Face Model Recommendation Chatbot",
    description="This chatbot recommends models from Hugging Face based on your query."
)

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