README.md

---
language:
  - code
license: apache-2.0
task_categories:
  - feature-extraction
  - sentence-similarity
tags:
  - code-search
  - hard-negatives
  - knowledge-distillation
  - contrastive-learning
  - sentence-transformers
  - colbert
pretty_name: "Owl Code Search Hard Negative Datasets (Pre-KD)"
size_categories:
  - 1M<n<10M
dataset_info:
  - config_name: documents_go
    features:
      - name: document_id
        dtype: string
      - name: document
        dtype: string
      - name: split
        dtype: string
    splits:
      - name: train
        num_bytes: 788961766
        num_examples: 1361475
    download_size: 234362060
    dataset_size: 788961766
  - config_name: documents_java
    features:
      - name: document_id
        dtype: string
      - name: document
        dtype: string
      - name: split
        dtype: string
    splits:
      - name: train
        num_bytes: 595749922
        num_examples: 1281018
    download_size: 157335988
    dataset_size: 595749922
  - config_name: documents_javascript
    features:
      - name: document_id
        dtype: string
      - name: document
        dtype: string
      - name: split
        dtype: string
    splits:
      - name: train
        num_bytes: 103608078
        num_examples: 129007
    download_size: 36381974
    dataset_size: 103608078
  - config_name: documents_php
    features:
      - name: document_id
        dtype: string
      - name: document
        dtype: string
      - name: split
        dtype: string
    splits:
      - name: train
        num_bytes: 223866454
        num_examples: 424463
    download_size: 63038942
    dataset_size: 223866454
  - config_name: documents_python
    features:
      - name: document_id
        dtype: string
      - name: document
        dtype: string
      - name: split
        dtype: string
    splits:
      - name: train
        num_bytes: 1076037918
        num_examples: 776900
    download_size: 335083048
    dataset_size: 1076037918
  - config_name: documents_ruby
    features:
      - name: document_id
        dtype: string
      - name: document
        dtype: string
      - name: split
        dtype: string
    splits:
      - name: train
        num_bytes: 63811504
        num_examples: 104899
    download_size: 15337714
    dataset_size: 63811504
  - config_name: documents_rust
    features:
      - name: document_id
        dtype: string
      - name: document
        dtype: string
      - name: split
        dtype: string
    splits:
      - name: train
        num_bytes: 239736870
        num_examples: 381521
    download_size: 76499102
    dataset_size: 239736870
  - config_name: documents_typescript
    features:
      - name: document_id
        dtype: string
      - name: document
        dtype: string
      - name: split
        dtype: string
    splits:
      - name: train
        num_bytes: 265760164
        num_examples: 328457
    download_size: 77031340
    dataset_size: 265760164
  - config_name: queries_go
    features:
      - name: query_id
        dtype: string
      - name: query
        dtype: string
      - name: split
        dtype: string
    splits:
      - name: train
        num_bytes: 244424686
        num_examples: 1361475
    download_size: 79920936
    dataset_size: 244424686
  - config_name: queries_java
    features:
      - name: query_id
        dtype: string
      - name: query
        dtype: string
      - name: split
        dtype: string
    splits:
      - name: train
        num_bytes: 392045510
        num_examples: 1281018
    download_size: 101975044
    dataset_size: 392045510
  - config_name: queries_javascript
    features:
      - name: query_id
        dtype: string
      - name: query
        dtype: string
      - name: split
        dtype: string
    splits:
      - name: train
        num_bytes: 40990466
        num_examples: 129007
    download_size: 14437070
    dataset_size: 40990466
  - config_name: queries_php
    features:
      - name: query_id
        dtype: string
      - name: query
        dtype: string
      - name: split
        dtype: string
    splits:
      - name: train
        num_bytes: 101841366
        num_examples: 424463
    download_size: 31103178
    dataset_size: 101841366
  - config_name: queries_python
    features:
      - name: query_id
        dtype: string
      - name: query
        dtype: string
      - name: split
        dtype: string
    splits:
      - name: train
        num_bytes: 333616210
        num_examples: 776900
    download_size: 102401118
    dataset_size: 333616210
  - config_name: queries_ruby
    features:
      - name: query_id
        dtype: string
      - name: query
        dtype: string
      - name: split
        dtype: string
    splits:
      - name: train
        num_bytes: 67053734
        num_examples: 104899
    download_size: 16664920
    dataset_size: 67053734
  - config_name: queries_rust
    features:
      - name: query_id
        dtype: string
      - name: query
        dtype: string
      - name: split
        dtype: string
    splits:
      - name: train
        num_bytes: 78330830
        num_examples: 381521
    download_size: 30435878
    dataset_size: 78330830
  - config_name: queries_typescript
    features:
      - name: query_id
        dtype: string
      - name: query
        dtype: string
      - name: split
        dtype: string
    splits:
      - name: train
        num_bytes: 96598646
        num_examples: 328457
    download_size: 30449400
    dataset_size: 96598646
  - config_name: scores_go
    features:
      - name: query_id
        dtype: string
      - name: document_ids
        sequence: string
      - name: scores
        sequence: float64
      - name: split
        dtype: string
    splits:
      - name: train
        num_bytes: 1049496498
        num_examples: 1361475
    download_size: 556865396
    dataset_size: 1049496498
  - config_name: scores_java
    features:
      - name: query_id
        dtype: string
      - name: document_ids
        sequence: string
      - name: scores
        sequence: float64
      - name: split
        dtype: string
    splits:
      - name: train
        num_bytes: 1072717102
        num_examples: 1281018
    download_size: 525421824
    dataset_size: 1072717102
  - config_name: scores_javascript
    features:
      - name: query_id
        dtype: string
      - name: document_ids
        sequence: string
      - name: scores
        sequence: float64
      - name: split
        dtype: string
    splits:
      - name: train
        num_bytes: 129963078
        num_examples: 129007
    download_size: 50592770
    dataset_size: 129963078
  - config_name: scores_php
    features:
      - name: query_id
        dtype: string
      - name: document_ids
        sequence: string
      - name: scores
        sequence: float64
      - name: split
        dtype: string
    splits:
      - name: train
        num_bytes: 335188010
        num_examples: 424463
    download_size: 172108596
    dataset_size: 335188010
  - config_name: scores_python
    features:
      - name: query_id
        dtype: string
      - name: document_ids
        sequence: string
      - name: scores
        sequence: float64
      - name: split
        dtype: string
    splits:
      - name: train
        num_bytes: 696396830
        num_examples: 776900
    download_size: 324380260
    dataset_size: 696396830
  - config_name: scores_ruby
    features:
      - name: query_id
        dtype: string
      - name: document_ids
        sequence: string
      - name: scores
        sequence: float64
      - name: split
        dtype: string
    splits:
      - name: train
        num_bytes: 83655522
        num_examples: 104899
    download_size: 40260810
    dataset_size: 83655522
  - config_name: scores_rust
    features:
      - name: query_id
        dtype: string
      - name: document_ids
        sequence: string
      - name: scores
        sequence: float64
      - name: split
        dtype: string
    splits:
      - name: train
        num_bytes: 314038494
        num_examples: 381521
    download_size: 163389490
    dataset_size: 314038494
  - config_name: scores_typescript
    features:
      - name: query_id
        dtype: string
      - name: document_ids
        sequence: string
      - name: scores
        sequence: float64
      - name: split
        dtype: string
    splits:
      - name: train
        num_bytes: 337032714
        num_examples: 328457
    download_size: 132654824
    dataset_size: 337032714
configs:
  - config_name: documents_go
    data_files:
      - split: train
        path: documents_go/train-*
  - config_name: documents_java
    data_files:
      - split: train
        path: documents_java/train-*
  - config_name: documents_javascript
    data_files:
      - split: train
        path: documents_javascript/train-*
  - config_name: documents_php
    data_files:
      - split: train
        path: documents_php/train-*
  - config_name: documents_python
    data_files:
      - split: train
        path: documents_python/train-*
  - config_name: documents_ruby
    data_files:
      - split: train
        path: documents_ruby/train-*
  - config_name: documents_rust
    data_files:
      - split: train
        path: documents_rust/train-*
  - config_name: documents_typescript
    data_files:
      - split: train
        path: documents_typescript/train-*
  - config_name: queries_go
    data_files:
      - split: train
        path: queries_go/train-*
  - config_name: queries_java
    data_files:
      - split: train
        path: queries_java/train-*
  - config_name: queries_javascript
    data_files:
      - split: train
        path: queries_javascript/train-*
  - config_name: queries_php
    data_files:
      - split: train
        path: queries_php/train-*
  - config_name: queries_python
    data_files:
      - split: train
        path: queries_python/train-*
  - config_name: queries_ruby
    data_files:
      - split: train
        path: queries_ruby/train-*
  - config_name: queries_rust
    data_files:
      - split: train
        path: queries_rust/train-*
  - config_name: queries_typescript
    data_files:
      - split: train
        path: queries_typescript/train-*
  - config_name: scores_go
    data_files:
      - split: train
        path: scores_go/train-*
  - config_name: scores_java
    data_files:
      - split: train
        path: scores_java/train-*
  - config_name: scores_javascript
    data_files:
      - split: train
        path: scores_javascript/train-*
  - config_name: scores_php
    data_files:
      - split: train
        path: scores_php/train-*
  - config_name: scores_python
    data_files:
      - split: train
        path: scores_python/train-*
  - config_name: scores_ruby
    data_files:
      - split: train
        path: scores_ruby/train-*
  - config_name: scores_rust
    data_files:
      - split: train
        path: scores_rust/train-*
  - config_name: scores_typescript
    data_files:
      - split: train
        path: scores_typescript/train-*
---

# Owl Code Search Hard Negative Datasets

Knowledge Distillation (KD) ベースのハードネガティブ付きコード検索データセットです。  
コード検索モデル[Shuu12121/CodeSearch-ModernBERT-Crow-v3-large-len1024-Plus](https://huggingface.co/Shuu12121/CodeSearch-ModernBERT-Crow-v3-large-len1024-Plus)を教師モデルとして、[各コメントと説明コメントのペアのデータセット](https://huggingface.co/collections/Shuu12121/codesearch-datasets)から各クエリに対する関数の類似度スコアを計算し、ハードネガティブ（正解に類似しているが不正解の文書）を付与しています。

## 概要

- **目的**: コード検索モデルの Contrastive Learning / Knowledge Distillation ファインチューニング
- **言語**: Go, Java, JavaScript, PHP, Python, Ruby, Rust, TypeScript（8言語）
- **総サンプル数**: 4,787,740
- **データサイズ**: 8.73 GB（展開後） / 3.37 GB（ダウンロード時）
- **フォーマット**: Per-language config 形式（`scores_{lang}`, `queries_{lang}`, `documents_{lang}`）

## データ構造

各言語ごとに 3 つの config が存在します：

### `queries_{lang}`

各クエリ（自然言語による検索文）を格納。

| カラム | 型 | 説明 |
|--------|------|------|
| `query_id` | `string` | クエリの一意識別子 |
| `query` | `string` | 自然言語のクエリテキスト（docstring / コメント） |
| `split` | `string` | 元データの分割情報 |

### `documents_{lang}`

各文書（ソースコード）を格納。

| カラム | 型 | 説明 |
|--------|------|------|
| `document_id` | `string` | 文書の一意識別子 |
| `document` | `string` | ソースコード本文 |
| `split` | `string` | 元データの分割情報 |

### `scores_{lang}`

教師モデルによる類似度スコアを格納。各クエリに対して、スコア順にソートされた文書 ID リストとスコアリストを保持。

| カラム | 型 | 説明 |
|--------|------|------|
| `query_id` | `string` | 対応するクエリの ID |
| `document_ids` | `list[string]` | スコア順にソートされた文書 ID のリスト |
| `scores` | `list[float64]` | 対応する類似度スコアのリスト |
| `split` | `string` | 元データの分割情報 |

> **スコアの解釈**:  
> - `scores[0]` / `document_ids[0]` が正例（実際のペアだったもの）
> - `score[0] = -1` は正解が上位32件に検索結果が含まれていなかった場合

## 言語別統計

| 言語 | クエリ数 | 文書数 | スコア数 |
|------|-------:|-------:|-------:|
| Go | 1,361,475 | 1,361,475 | 1,361,475 |
| Java | 1,281,018 | 1,281,018 | 1,281,018 |
| JavaScript | 129,007 | 129,007 | 129,007 |
| PHP | 424,463 | 424,463 | 424,463 |
| Python | 776,900 | 776,900 | 776,900 |
| Ruby | 104,899 | 104,899 | 104,899 |
| Rust | 381,521 | 381,521 | 381,521 |
| TypeScript | 328,457 | 328,457 | 328,457 |
| **合計** | **4,787,740** | **4,787,740** | **4,787,740** |

## 注意点
全データをメモリに載せようとするとOOMになる可能性があります！！


## 使い方

### 基本的な読み込み

```python
from datasets import load_dataset

# Python の scores を読み込む
scores = load_dataset(
    "Shuu12121/owl_code_search_hard_negative_datasets-Pre_kd",
    name="scores_python",
    split="train",
)

# Python の queries を読み込む
queries = load_dataset(
    "Shuu12121/owl_code_search_hard_negative_datasets-Pre_kd",
    name="queries_python",
    split="train",
)

# Python の documents を読み込む
documents = load_dataset(
    "Shuu12121/owl_code_search_hard_negative_datasets-Pre_kd",
    name="documents_python",
    split="train",
)
```
### ハードネガティブの抽出
```python
# クエリ・文書テキストの辞書を構築
query_texts = dict(zip(queries["query_id"], queries["query"]))
doc_texts = dict(zip(documents["document_id"], documents["document"]))

# 閾値の設定
nv_threshold = 0.99  # positive スコアの 99% 未満をネガティブとする

# 1 サンプルの処理例
sample = scores[0]
query_text = query_texts[sample["query_id"]]
positive_doc = doc_texts[sample["document_ids"][0]]  # scores[0] が正例
positive_score = sample["scores"][0]

hard_negatives = []
for doc_id, score in zip(sample["document_ids"][1:], sample["scores"][1:]):
    if score < nv_threshold * positive_score and score != -1:
        hard_negatives.append(doc_texts[doc_id])

print(f"Query: {query_text[:100]}...")
print(f"Positive: {positive_doc[:100]}...")
print(f"Hard negatives: {len(hard_negatives)}")
```
##  作成に使用されたプログラム
[リポジトリはこちら](https://github.com/Shun0212/hard-negatives-ranking-datasets-maker)