BranLiu/cr_taxonomy_dataset

Code Review Comment Taxonomy — Dataset Collection & Preprocessing

Project Purpose

本项目旨在收集市面上最全的 Code Review Comment 数据样本，应用分类方法，对目前 Code Review 的主要类型进行分类（empirical taxonomy）。

目录结构

├── README.md
├── data/
│   ├── merged.jsonl                  # 1,142,862 rows (25 个数据集融合后，未去重)
│   ├── final.jsonl                   # 747,352 rows (精确+语义去重后) ← 最终数据
│   ├── dedup_exact_report.json       # 精确去重报告
│   ├── dedup_semantic_report.json    # 语义去重报告
│   └── stats.json                    # 统计信息
├── scripts/
│   ├── schema.py                     # 统一 schema 定义
│   ├── convert_*.py                  # 各数据集转换脚本 (25 个)
│   ├── merge_all.py                  # 数据集融合
│   ├── dedup_exact.py                # 精确去重
│   ├── dedup_semantic.py             # 语义去重
│   ├── compute_stats.py              # 统计计算
│   └── run_pipeline.py               # 一键运行

---

已收集数据集清单

Core Corpus（训练语料，10 个数据集，712,140 条）

#	数据集	论文/来源	原始格式	转换后行数	Split	关键特点
1	CodeReviewer	Li et al., "Automating Code Review Activities by Large-Scale Pre-Training", FSE 2022 · HF / Zenodo	JSONL (oldf, patch, msg, id, y)	129,327	train/valid/test	事实标准训练 corpus
2	CuREV	Sghaier et al., "CuREV: Constructing a Benchmark for Evaluating the Quality of Code Review Comments", MSR 2025 · Zenodo	JSONL + quality scores	5,000	test	质量评估标注 (Type/Nature/Civility)
3	ContextCRBench	2025 · GitHub	单 JSON per review (68,938 files)	58,096	test	唯一带完整 code context
4	Gerrit	Yang et al., "Mining Review Repositories to Evaluate People, Process and Product", MSR 2016 · Website	MySQL dump (t_history, t_revision)	16,429	test	OpenStack/LibreOffice/Eclipse/GerritHub
5	Tufano	Tufano et al., "Using Pre-Trained Models to Automate Code Review Comment Generation", ICSE 2022 · GitHub	TSV (code-to-comment)	167,797	train/test/valid	早期 baseline，Java
6	ContextualCodeReview	Muttakin, 2024 · HF	JSON (comment, method_body, target_code)	60,845	test	Java，带方法级上下文
7	CROP	Thongtanunam et al., "Revisiting Code Review Mediation: Is It All About the Reviewer?", MSR 2015 (Zenodo 2020) · Zenodo	discussion.zip (txt files) + metadata.zip (CSV)	86,737	test	11 个 Gerrit 项目，含 inline comment
8	SAILCRCoverage	McIntosh et al., "The Impact of Code Review Coverage and Code Review Participation on Software Quality", MSR 2014 · GitHub	MySQL dump (Comment/Review/Message 表)	165,168	test	Qt/VTK Gerrit，无 diff_hunk
9	CRC-Py	Icoz, 2025 · GitHub	JSON	13,726	train/test	5 类 13 子类
10	CleanCodeReview	Petrova et al., Zenodo 2024 · Zenodo	JSONL (message, group, label)	10,022	train/test	16 类 / 5 超组

Benchmark（评测/标注集，15 个数据集，35,212 条）

#	数据集	论文/来源	原始格式	转换后行数	Split	分类标签
11	CodeReviewQA	HF	JSONL	899	test	无
12	CodeFuse-CR-Bench	Yang et al., arXiv 2509.14856	Parquet	12,881	test	无
13	SeRe	Shen et al., "SeRe: A Security-Related Code Review Comment Generation Approach", ICSE 2026 · arXiv 2601.01042	JSONL	6,729	test	Security 相关标注
14	CR-ESEM23	Turzo et al., "Automated Classification of Code Review Comments Using Deep Learning", ESEM 2023 · GitHub	Excel (44 cols, labeled)	1,694	test	category + comment_group + usefulness
15	AACR-Bench	Zhang et al., "AACR-Bench: A Benchmark for Automated AI-Code-Review", ICML 2026 (arXiv 2601.19494) · GitHub	JSON	2,133	test	positive/negative
16	RevHelper	Rahman et al., "RevHelper: A Tool to Support Code Reviewers", MSR 2017 · GitHub	TSV	363	test	useful/non-useful
17	ConfusionCR	Ebert et al., "Confusion in Code Reviews: Reasons, Impacts, and Perceptions", TSE · GitHub	XLS	2,714	test	confusion/non-confusion
18	TooNoisy	Li et al., arXiv 2502.02757 · Zenodo	JSONL	270	test	Valid/Noisy
19	c-CRAB	Rossi et al., "c-CRAB: A Benchmark for Evaluating Automated Code Review Comment Filtering and Resolution", ICSE 2025 · GitHub	JSONL	595	test	无
20	SWRench	GitHub	JSONL	910	test	change_type (如 F.2 Logic)
21	EvaCRC	Li et al., "EvaCRC: Evaluating Code Review Comments", ESEC/FSE 2023 · Zenodo	Excel	2,910	test	Emotion/Question/Evaluation/Suggestion + Grade
22	CR Smell	arXiv 2604.23667 · Figshare	CSV	439	test	9 类 (6 smell + 3 intent)，有 diff_hunk
23	TestingMeetsCR	Moritz et al., "When Testing Meets Code Review", ICSE 2018 · Zenodo	Excel (CategorizationOfComments.xlsx)	581	test	Code improvement/Defect/Knowledge Transfer/Misc + 细粒度子类
24	DesignDiscussions	Uchoa et al., "An Empirical Study of Design Discussions In Code Review", ESEM 2018 · GitHub	XLS	2,359	test	is_design + design_label
25	ArchAwareness	Paixao et al., "Are Developers Aware of the Architectural Impact of Their Changes?", SANER 2017 · Website	CSV + review txt files	836	test	变更类型 + 架构感知 (author_aware/comments_aware/not_aware)

---

预处理流程

Step 1: 原始格式 → 统一 JSONL

每个数据集由独立的 scripts/convert_<dataset>.py 转换。统一 Schema：

{
  "source_datasets": "list[dict], 每项 {\"dataset\": str, \"split\": str|list[str]}，去重合并后该样本的所有来源",
  "repo": "str, owner/repo",
  "commit_id": "str, commit hash",
  "pr_number": "int or null",
  "comment_id": "str, 数据集内唯一标识",
  "file_path": "str or null",
  "diff_hunk": "str, unified diff / patch",
  "comment": "str, review comment text",
  "lang": "str or null, 标准化后",
  "created_at": "str or null",
  "extra_fields": "dict, 数据集特有字段",
  "in_benchmark": "bool"
}

语言标准化：C# / csharp → c_sharp; C++ → cpp; Golang → go; 其余统一小写。

Split 规则：

• Benchmark 数据集：全部标为 test
• Corpus 数据集有原始 split 信息：保留原始 split（train/test/valid）
• Corpus 数据集无 split 信息：全部标为 test

注意：少数行因跨数据集去重合并，split 字段可能为 list（如 ["test", "valid"] 或 ["test", "train"]），共 252 行，使用时需注意潜在的 train/test 泄露。

Step 2: 融合 (merge_all.py)

将 data/raw/*.jsonl (10 corpus) + data/benchmark/*.jsonl (15 benchmark) 纯拼接合并。

• 为每行构建 _unique_key = (repo, commit_id, comment_id) 和 _content_hash = SHA256(normalized_diff + normalized_comment)
• 标记 in_benchmark 字段
• 语言名标准化

结果：25 个数据集 → 1,142,862 rows

Step 3: 精确去重 (dedup_exact.py)

1. Unique key 去重: 仅对 repo、commit_id、comment_id 三字段均存在的行，按 (repo, commit_id, comment_id) 去重，同 key 只保留一条，来源合并到 source_datasets
2. Content hash 去重: SHA256(normalized_diff + normalized_comment) 相同只保留一条，来源合并到 source_datasets；若 _content_hash 为空则重新计算

结果：1,142,862 → 747,356（移除 395,506 条，其中 key 重复 1,344 条，hash 重复 394,162 条）

Step 4: 语义去重 (dedup_semantic.py)

跨数据集近似重复检测：

1. 按 (repo, comment前50字符) 分桶
2. 仅在跨数据集的桶内做比较
3. 计算 Jaccard token 相似度，阈值 comment ≥ 0.85 AND diff ≥ 0.9 → 语义重复

结果：747,356 → 747,352（移除 4 条）

Step 5: 统计 (compute_stats.py)

指标	值
总行数	747,352
Corpus	712,140
Benchmark	35,212
唯一 repo	691

按数据集分布：

数据集	行数	占比	类型
Tufano	167,797	22.5%	corpus
SAILCRCoverage	165,168	22.1%	corpus
CodeReviewer	129,327	17.3%	corpus
CROP	86,737	11.6%	corpus
ContextualCodeReview	60,845	8.2%	corpus
ContextCRBench	58,096	7.8%	corpus
Gerrit	16,429	2.2%	corpus
CRC-Py	13,726	1.8%	corpus
CodeFuse-CR-Bench	12,881	1.7%	benchmark
CleanCodeReview	10,022	1.3%	corpus
SeRe	6,731	0.9%	benchmark
CuREV	5,000	0.7%	corpus
EvaCRC	2,910	0.4%	benchmark
ConfusionCR	2,714	0.4%	benchmark
DesignDiscussions	2,359	0.3%	benchmark
AACR-Bench	2,133	0.3%	benchmark
CR-ESEM23	1,694	0.2%	benchmark
SWRench	910	0.1%	benchmark
CodeReviewQA	899	0.1%	benchmark
ArchAwareness	836	0.1%	benchmark
c-CRAB	595	0.1%	benchmark
TestingMeetsCR	581	0.1%	benchmark
CRSmell	439	0.1%	benchmark
RevHelper	363	<0.1%	benchmark
TooNoisy	270	<0.1%	benchmark

按 Split 分布：

Split	行数	占比
test	459,952	61.3%
train	262,648	35.1%
valid	26,862	3.6%

按语言分布：

语言	行数	占比
unknown	424,263	56.9%
java	234,476	31.4%
python	37,881	5.1%
go	14,490	1.9%
c_sharp	10,962	1.5%
cpp	7,451	1.0%
c	7,166	1.0%
rust	4,593	0.6%
javascript	2,504	0.3%
typescript	2,356	0.3%
ruby	1,054	0.1%
php	156	<0.1%