app.py

# CodeSearch-ModernBERT-Owl Demo Space using CodeSearchNet Dataset
import gradio as gr
import torch
import random
from sentence_transformers import SentenceTransformer, util
from datasets import load_dataset
from spaces import GPU

# --- Load model ---
model = SentenceTransformer("Shuu12121/CodeSearch-ModernBERT-Owl")
model.eval()

# --- Load CodeSearchNet dataset (test split only) ---
dataset = load_dataset("code_x_glue_tc_nl_code_search_adv", trust_remote_code=True, split="test")

# --- Query & Candidate Generator ---
def get_random_query(seed: int = 42):
    random.seed(seed)
    idx = random.randint(0, len(dataset) - 1)
    sample = dataset[idx]
    return sample["code"], sample["docstring"]

@GPU
def code_search_demo(seed: int):
    code_str, doc_str = get_random_query(seed)
    query_emb = model.encode(doc_str, convert_to_tensor=True)

    # ランダムに10件取得
    candidates = dataset.shuffle(seed=seed).select(range(10))
    candidate_codes = [c["code"] for c in candidates]
    candidate_embeddings = model.encode(candidate_codes, convert_to_tensor=True)

    # 類似度スコア算出
    cos_scores = util.cos_sim(query_emb, candidate_embeddings)[0]
    results = sorted(zip(candidate_codes, cos_scores), key=lambda x: x[1], reverse=True)

    # 結果出力
    output = f"### 🔍 Query Docstring\n\n{doc_str}\n\n"
    output += "## 🏆 Top Matches:\n"
    medals = ["🥇", "🥈", "🥉"] + [f"#{i+1}" for i in range(3, len(results))]
    for i, (code, score) in enumerate(results):
        label = medals[i] if i < len(medals) else f"#{i+1}"
        output += f"\n**{label}** - Similarity: {score.item():.4f}\n\n```python\n{code.strip()[:1000]}\n```\n"

    return output

# --- Gradio UI ---
demo = gr.Interface(
    fn=code_search_demo,
    inputs=gr.Slider(0, 100000, value=42, step=1, label="Random Seed"),
    outputs=gr.Markdown(label="Search Result"),
    title="🔎 CodeSearch-ModernBERT-Owl Demo",
    description="docstring から類似 Python 関数を検索（CodeXGlue + ModernBERT-Owl）"
)

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