the-stack-go / stackgo /README.md
s0duku's picture
Add generic end-to-end dataset builder and smoke-tested tooling
370fed4 verified
|
Raw History Blame Contribute Delete
9.22 kB

stackgo

stackgo 是 The Stack v2 代码数据流水线和仓库级上下文样本工具。它把数据获取、星数区间筛选、正文下载、多区间合并导出和样本生成拆成可复现的阶段,并将 stackbuild 作为同一项目中的仓库扫描与样本生成库。

流程

  1. 下载 The Stack v2 的 repository-grouped file ID 索引(不含源码正文)。
  2. 从索引物化任意严格星数区间,例如 star_events_count > 10 或 1 < star_events_count < 11。
  3. 从 Software Heritage 按 blob_id 下载并 SHA-1 校验正文。
  4. 将一个或多个正文下载目录合并成统一布局,并按 (shard, blob_id) 去重。
  5. 将每个仓库快照传入 stackbuild,使用固定 DeepSeek V4 Flash Base tokenizer 和原生 repo/file 边界,生成真实 8K token 预算的 ZSTD Parquet 依赖上下文样本。

星数是索引快照中的 star_events_count,不是实时 GitHub star;> 10 不包含等于 10。索引没有 commit、issue 或 PR 历史。正文下载遵守 The Stack 和 Software Heritage 的访问条款,并保留原始许可信息。

上游说明:The Stack v2 IDs 和 官方正文下载流程。批量访问 Software Heritage 前应遵守其条款并按上游要求联系 datasets@softwareheritage.org;匿名 S3 可访问不代表免除条款。完整数据预估约 143 GiB 原始索引长度,建议为 Parquet、状态库和临时文件预留至少约 200 GiB。

安装

cd stackgo
python -m venv .venv
.venv/bin/pip install -e '.[test]'

获取数据

stackgo-download-index --output-dir /mnt/data/the-stack-v2-train-full-ids
stackgo-filter-star10 \
  --input-dir /mnt/data/the-stack-v2-train-full-ids \
  --output-dir /mnt/data/the-stack-v2-train-full-ids-star-gt10
stackgo-download-contents \
  --index-dir /mnt/data/the-stack-v2-train-full-ids-star-gt10 \
  --output-dir /mnt/data/the-stack-v2-contents-star-gt10 \
  --already-filtered --workers 16
stackgo-export \
  --source /mnt/data/the-stack-v2-contents-star-gt10 \
  --output /mnt/data/the-stack-go

任意区间筛选使用 stackgo-filter-stars;正文下载器也支持相同的严格上下界:

stackgo-filter-stars \
  --input-dir /mnt/data/the-stack-v2-train-full-ids \
  --output-dir /mnt/data/the-stack-v2-train-full-ids-star-2-10 \
  --min-stars-exclusive 1 --max-stars-exclusive 11

stackgo-download-contents \
  --index-dir /mnt/data/the-stack-v2-train-full-ids-star-2-10 \
  --output-dir /mnt/data/the-stack-v2-contents-star-2-10 \
  --already-filtered --workers 16

构建 star_events_count > 1

如果已有 >10 数据,可增量下载 2–9 和精确 10,再合并,避免重下 >10 正文:

# 精确 10 星:9 < star_events_count < 11
stackgo-download-contents \
  --index-dir /mnt/data/the-stack-v2-train-full-ids \
  --output-dir /mnt/data/the-stack-v2-contents-star-10 \
  --min-stars-exclusive 9 --max-stars-exclusive 11 --workers 8

stackgo-export \
  --source /mnt/data/the-stack-v2-contents-star-gt10 \
  --source /mnt/data/the-stack-v2-contents-star-2-9 \
  --source /mnt/data/the-stack-v2-contents-star-10 \
  --output /mnt/data/the-stack-go-ge2 \
  --star-events-count-exclusive 1 \
  --expected-repositories 3246224 \
  --expected-files 81112563

stackgo-build-samples --root /mnt/data/the-stack-go-ge2 --jobs 4

合并导出会检查各来源是否下载完成、星数区间是否连续、仓库是否重复,并校验最终仓库数和文件引用数。stackgo-export --source 可重复传入;只有一个来源时仍保持原有的硬链接导出模式。

也可以用一个可恢复入口完成“等待下载、合并校验、样本生成”:

stackgo-build-ge2 --wait --jobs 4

默认读取 the-stack-v2-contents-star-gt10、the-stack-v2-contents-star-2-9 和 the-stack-v2-contents-star-10,输出 the-stack-go-ge2/。合并阶段中断后可重跑,已完成且来源身份一致的分片会跳过;样本阶段同样会跳过已完成分片。

任意 star 规则端到端构建

stackgo-build-dataset 从原始索引开始串联索引下载(缺失时)、区间过滤、正文下载、有限/无限区间导出,并可继续生成样本:

stackgo-build-dataset \
  --min-stars-exclusive 5 \
  --max-stars-exclusive 11 \
  --output /mnt/data/the-stack-go-stars-6-10 \
  --workers 64 \
  --with-samples --jobs 32

不传 --max-stars-exclusive 表示只有严格下界,例如 --min-stars-exclusive 1 表示 >1。已有索引时可加 --skip-index-download 复用 --index-dir。冒烟测试可以限制源分片和正文请求:

stackgo-build-dataset \
  --min-stars-exclusive 10000 \
  --max-stars-exclusive 20000 \
  --output /tmp/stackgo-smoke \
  --skip-index-download --max-source-shards 1 --max-files 3 \
  --with-samples --jobs 1 --sample-max-repositories 1

The index is gated on Hugging Face: accept its dataset terms and configure HF_TOKEN or hf auth login first. STACKGO_DATA_ROOT changes defaults when explicit paths are omitted.

所有步骤都可安全重启。正文下载器按 blob 去重、原子发布 Parquet、记录 SQLite 状态,并验证正文 SHA-1。--status 查看状态,--index-only 只建立正文下载队列;Software Heritage 暂不可用的对象可在确认来源后使用 stackgo-recover-contents 尝试精确 revision 的 GitHub fallback。

stackgo-export 将已下载数据拆成两个目录:repo-meta/<shard>/repositories.parquet 保存仓库、文件路径、revision 和许可等元数据;repo-contents/<shard>/*.parquet 保存 blob_id、正文和长度。两者通过 shard 名和 blob_id 关联。默认使用硬链接,不额外占用正文空间;跨文件系统时使用 --mode copy,或用 --mode symlink 建立绝对链接。输出的 manifest.json 记录来源 manifest 哈希、布局和统计。

导出目录会自动生成根目录、repo-meta/ 和 repo-contents/ 三份 README,说明字段和关联方式,适合直接作为 Hugging Face dataset repository 的数据说明。项目代码本身就在 the-stack-go/stackgo/ 中维护,不由导出器复制或刷新。

生成样本

from stackbuild import BuildConfig, build_samples, write_parquet

samples = build_samples(
    repo,
    config=BuildConfig(
        max_tokens=8192,
        target_context=True,
        max_targets_per_repo=32,
        max_dependency_depth=3,
    ),
)
write_parquet(samples, "repository-samples.parquet")

默认 tokenizer 固定为 deepseek-ai/DeepSeek-V4-Flash-Base revision 8855555deef230a27a21a8d6f294b7b7497759b6。优先读取同级 ../tokenizers/deepseek-v4-flash-base/,也可用 STACKBUILD_TOKENIZER_PATH 指定本地目录。 全量生成使用本地 tokenizer,不会在扫描中下载模型或 tokenizer:

python -m stackgo.scripts.build_stack_samples --root /mnt/data/the-stack-go --jobs 8

输出为 samples/train-xxxxx-of-00064.parquet,使用 ZSTD 压缩和 2048 行 row group。 manifest.json 记录 tokenizer revision、SHA-256 和预算;安全重启会跳过配置一致且已完成的分片。 中断的 .parquet.part 会重新生成。默认并行构建 8 个源 shard,每处理 100 个仓库打印进度; 内存较小的机器可用 --jobs 1 降低并发。

每个仓库在独立的可杀死 runner 中构建。默认单仓库超过 1800 秒时, 只终止并降级当前仓库到 regex,其他仓库和后续分片继续优先使用 tree-sitter。

dependency 样本覆盖仓库全部文件;target_context 样本优先放置直接和传递依赖,再放 README、manifest 和构建文件。二进制、vendor 和 generated 文件默认只保留元数据。完整 API 说明见 stackbuild/README.md。

读取样本

样本是标准 Parquet 文件,推荐流式读取。仓库提供了可运行示例:

PYTHONPATH=. python examples/read_samples.py ../samples/train-*.parquet --limit 3
PYTHONPATH=. python examples/read_samples.py ../samples \
  --repo owner/project --kind dependency --limit 5

PyArrow 流式读取:

import json
import pyarrow.dataset as ds

dataset = ds.dataset("samples", format="parquet")
scanner = dataset.scanner(
    columns=["sample_id", "repo_id", "kind", "text", "files", "token_count", "metadata_json"],
    filter=ds.field("kind") == "dependency",
    batch_size=512,
)
for batch in scanner.to_batches():
    for row in batch.to_pylist():
        print(row["repo_id"], row["token_count"])
        print([item["path"] for item in row["files"]])
        print(json.loads(row["metadata_json"]))

完整示例见 examples/README.md,包括 Hugging Face datasets 流式加载、按仓库/kind 过滤和文件 provenance 读取。

测试

PYTHONPATH=. pytest -q

数据目录不放在 Git 工作区;用 manifest、revision、过滤条件和生成配置记录每次数据版本。