You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

Public LLM Data Pool

本仓库维护公开语料的上游引用、revision、许可入口、采样账本和获取说明,并托管部分带来源记录的压缩样本。

数据规模

统计项 数量与口径
语料池目标 10T tokens,按使用方 tokenizer、清洗及跨源去重后统计
最大单一上游报告量 FineWeb 固定快照的上游报告量为 15T GPT-2 tokens;这是外部来源报告,不是本仓库存储量,也不是已验证的中英唯一 token 总量。
本仓库托管的正文 116.89M 代理 tokens / 136.86MB 压缩文件
预训练样本镜像 42,532 篇 / 61.45M 代理 tokens
SFT 样本镜像 2,550 条 / 55.44M 代理 tokens
已审核可训练 token 0M

注意:上游报告量、派生子集、镜像 pilot 和仓库正文不可重复相加。仓库代理 token 使用项目 v1 BPE-8K,仅描述样本规模,不代表生产计量或 train-ready 数量。FineWeb 的报告量使用 GPT-2 tokenizer;FineWeb-2 中文体量以 word 数/bytes 报告,不是 token 数;FinePDFs 的总量覆盖多个语言,不能全部归到中文和英文。当前跨来源、跨 tokenizer 的唯一可用 token 数尚未核定。

预训练来源目录

来源 语言 / 用途 上游报告量 本仓库样本 revision 状态
fineweb_en en / web_backbone 15T tokens (GPT-2; 上游报告) 981 篇 / 0.85M 9bb295ddab0e referenced_candidates_not_acquired
fineweb_edu_en en / education_bucket 1.3T tokens (上游报告) 946 篇 / 1.20M 87f09149ef47 referenced_candidates_not_acquired
fineweb2_zh zh / web_backbone 543.5B words, 2.42TB UTF-8 (非 token) 949 篇 / 1.26M af9c13333eb9 referenced_candidates_not_acquired
ultrafineweb zh_en / quality_bucket EN 1000.0B; ZH 120.0B — 02c85641e3d1 referenced_candidates_not_acquired
finepdfs zh_en / pdf_knowledge 2.92T tokens (all langs); EN 1190.7B; ZH 33.0B — 220bac3acbf0 referenced_candidates_not_acquired
finepdfs_edu zh_en / education_bucket ≥350B tokens (all langs) — 9cfabe2127fa referenced_candidates_not_acquired
dclm en / alternative_web_selection 未核定 token 总量 39,656 篇 / 58.13M a3b142c183ae referenced_candidates_not_acquired
industry2 zh_en / industry_knowledge 未核定 token 总量 — 249ea52f1ee4 pending_access_terms_or_schema_review
cci3_hq zh / quality_web 未核定 token 总量 — d6f3aa30cebf pending_access_terms_or_schema_review
chinesewebtext2 zh / labeled_web 未核定 token 总量 — 9b647306646e referenced_candidates_not_acquired
opencsg_edu_zh zh / education_bucket 未核定 token 总量 — a5b574efa48b pending_access_terms_or_schema_review
nemotron_cc_v2 primarily_en_with_multilingual_synthetic / web_and_existing_synthetic 未核定 token 总量 — 2669787c66d1 pending_access_terms_or_schema_review
finemath en / math 未核定 token 总量 — e92b25a61673 referenced_candidates_not_acquired
stack_edu code / educational_code 未核定 token 总量 — eeec5caac5cc pending_access_terms_or_schema_review

上游许可、访问门槛和数据卡以各来源目录为准。教育质量子集、Common Crawl 替代筛选以及其他派生数据不作为独立新增量。

SFT 来源目录

来源 语言 / 类别 上游行数 镜像样本 声明许可 获取方式
toolace zh,en / bilingual_function_calling 11,300 1,500 行 / 1.52M apache-2.0 mirrored_bounded_pilot; upstream remains canonical full source
orca_agentinstruct en / synthetic_instruction_and_context_qa 1,046,410 650 行 / 1.45M cdla-permissive-2.0 mirrored_bounded_pilot; upstream remains canonical full source
xyz_aquila_en en / search_agent_trajectory 5,000 200 行 / 27.86M apache-2.0 mirrored_bounded_pilot; upstream remains canonical full source
xyz_aquila_zh zh / search_agent_trajectory 2,000 200 行 / 24.61M apache-2.0 mirrored_bounded_pilot; upstream remains canonical full source
code_feedback_pending en / code generation and execution feedback — — apache-2.0 仅登记来源

镜像样本是有界窗口候选,使用项目 v1 BPE-8K 估算 token,仅用于规模参考;源间去重、内容质量、评测污染和再分发审核未完成。文件保留上游来源、revision 与行号等字段;许可和归属以各自上游卡片为准。

获取说明

数据源、固定 revision、许可链接和命令见 数据池 Agent Skill。目录 JSON 位于 catalog/。

把数据池 skill 安装到当前项目的标准 Agent Skills 路径:

hf download hqsiswiliam/data_pool .agents/skills/data-pool/SKILL.md --repo-type dataset --revision main --local-dir .

每次添加、移除或扩大数据样本时,先更新对应 catalog 与 catalog/accounting.json,再运行:

python scripts/update_readme.py

提交前核对 README 中的来源、行数、token 口径、实际托管体积和许可链接。不要把上游报告量称作本仓库已收集量。

计量与状态字段

  • referenced_candidates_not_acquired: 仅有上游引用,可直接从原始 dataset 获取。
  • pending_access_terms_or_schema_review: 访问、许可或字段尚未核实。
  • train_ready_tokens: 只有完成质量、许可证、去污染与分组切分审核后才可大于零。
  • M/B/T: 十进制 token 数量级;分来源 tokenizer 及统计范围见各 manifest。
Downloads last month
169