hasankursun/codesearchnet-hard-negatives

CodeSearchNet Hard Negatives (Filtered) by Lumees AI

Dataset Summary

This dataset is a processed version of the CodeSearchNet dataset, enhanced with Hard Negative Mining to facilitate the training of state-of-the-art code retrieval models.

It was created by Lumees AI to improve the ability of embedding models to distinguish between syntactically similar but functionally different code snippets.

• Developer: Lumees AI
• Authors: Hasan Kurşun, Kerem Berkay Yanık
• Contact: hello@lumees.io
• Source Data: CodeSearchNet (Train split)
• Total Samples: ~1.88M triplets/tuples

Dataset Structure

The dataset is provided in .jsonl format. Each line represents a training sample containing a natural language query, the positive ground truth code, and a list of mined hard negatives.

Data Fields

• query (string): The natural language docstring/description of the function.
• pos (string): The positive (ground truth) code snippet.
• neg (list of strings): A list of hard negative code snippets (semantically similar to the query but incorrect).
• scores (list of floats): The cosine similarity scores of the negative candidates against the query (computed by the mining model).

Example Instance

{
  "query": "CommentsView sub-view (will be used recursively)",
  "pos": "function ThreadBranchView(vm) { ... }",
  "neg": [
    "function CommentReplyView(vm, comment) { ... }",
    "public function viewAction() { ... }"
  ],
  "scores": [0.7502, 0.7481]
}

Methodology & Creation

Source Model

The mining process utilized Alibaba-NLP/gte-multilingual-base, a high-performance embedding model, to generate vector representations for both queries and code.

Mining Process

The dataset was constructed using a dense retrieval approach on the entire CodeSearchNet training corpus across 6 languages (Python, Java, Go, PHP, Ruby, JavaScript).

1. Embedding: All code snippets in the corpus were encoded into dense vectors.
2. Retrieval: For every query, we retrieved the top 50 semantic candidates from the corpus using GPU-accelerated Matrix Multiplication.
3. Filtration:
   • Self-Exclusion: The positive ground truth was removed from results.
   • Duplicate Removal: Exact string duplicates of the positive code were removed.
   • Score Thresholding:
     • Max Similarity (0.95): Candidates with scores above 0.95 were discarded to avoid False Negatives (valid code that is too similar to the ground truth).
     • Min Similarity (0.35): Candidates with scores below 0.35 were discarded to ensure the negatives are "hard" enough to be useful for training (avoiding easy negatives).
4. Selection: Up to the top 12 valid hard negatives were selected for each query.

Intended Use

This dataset is optimized for:

• Contrastive Learning: Fine-tuning embedding models using losses like MultipleNegativesRankingLoss or TripletLoss.
• Code Retrieval: Improving search relevance in IDEs or code search engines.
• Cross-Lingual Alignment: The dataset includes cross-lingual negatives (e.g., a Python query retrieving similar PHP code), helping models learn language-agnostic semantic features.

Licensing

This dataset adheres to the licensing terms of the original CodeSearchNet dataset (MIT/Permissive). Users should verify specific licensing requirements for individual code snippets if used for commercial code generation.

Citation

If you use this dataset, please cite Lumees AI and the original CodeSearchNet paper:

@misc{lumees2025hardnegatives,
  author = {Hasan KURŞUN, Kerem Berkay YANIK},
  title = {CodeSearchNet Hard Negatives (Filtered)},
  year = {2025},
  publisher = {Lumees AI},
  howpublished = {\url{[https://lumees.io](https://lumees.io)}},
  email = {hello@lumees.io}
}

@article{husain2019codesearchnet,
  title={CodeSearchNet Challenge: Evaluating the State of Semantic Code Search},
  author={Husain, Hamel and Wu, Ho-Hsiang and Gazit, Tiferet and Allamanis, Miltiadis and Brockschmidt, Marc},
  journal={arXiv preprint arXiv:1909.09436},
  year={2019}
}