Upload _script_for_eval.py

_script_for_eval.py

@@ -7,15 +7,15 @@ import datetime
 import subprocess
 import argparse
 import re
+import multiprocessing
+import numpy as np
 from openai import OpenAI
 from openai import OpenAIError
 from tqdm import tqdm
 from functools import partial
-import multiprocessing
 from datasets import load_dataset
-from sklearn.feature_extraction.text import TfidfVectorizer
+from sentence_transformers import SentenceTransformer, CrossEncoder
 from sklearn.metrics.pairwise import cosine_similarity
-import numpy as np
 
 client = OpenAI()
 
@@ -41,17 +41,36 @@ def fetch_dataset_examples(prompt, num_examples=0, use_similarity=False):
     dataset = load_dataset("patched-codes/synth-vuln-fixes", split="train")
     
     if use_similarity:
+        # Load a lightweight model for initial retrieval
+        retrieval_model = SentenceTransformer('all-MiniLM-L6-v2')
+        
+        # Load the cross-encoder model for reranking
+        rerank_model = CrossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2')
+        
+        # Extract user messages
         user_messages = [
             next(msg['content'] for msg in item['messages'] if msg['role'] == 'user')
             for item in dataset
         ]
         
-        vectorizer = TfidfVectorizer().fit(user_messages + [prompt])
-        user_vectors = vectorizer.transform(user_messages)
-        prompt_vector = vectorizer.transform([prompt])
+        # Encode the prompt and user messages for initial retrieval
+        prompt_embedding = retrieval_model.encode(prompt, convert_to_tensor=False)
+        corpus_embeddings = retrieval_model.encode(user_messages, convert_to_tensor=False, show_progress_bar=True)
+        
+        # Perform initial retrieval
+        similarities = cosine_similarity([prompt_embedding], corpus_embeddings)[0]
+        top_k = min(100, len(dataset))
+        top_indices = similarities.argsort()[-top_k:][::-1]
+        
+        # Prepare pairs for reranking
+        rerank_pairs = [[prompt, user_messages[idx]] for idx in top_indices]
+        
+        # Rerank using the cross-encoder model
+        rerank_scores = rerank_model.predict(rerank_pairs)
         
-        similarities = cosine_similarity(prompt_vector, user_vectors)[0]
-        top_indices = np.argsort(similarities)[-num_examples:][::-1]
+        # Sort by reranked score and select top examples
+        reranked_indices = [top_indices[i] for i in np.argsort(rerank_scores)[::-1][:num_examples]]
+        top_indices = reranked_indices
     else:
         top_indices = np.random.choice(len(dataset), num_examples, replace=False)