The dataset viewer should be available soon. Please retry later.
Agentic Env Index
An index of materials for building agent environments: not a list of environments, and no content is copied here.
It covers 958,330,115 public files and records (4.02 PB) from 88 sources. Each row says where one item can be downloaded, how large it is, which hashes verify it and under which license it is offered. The files themselves stay at their sources.
Use
This repository gets new commits every day. Pin one commit so that the same query gives the same rows later. First note the current commit:
curl -s https://huggingface.co/api/datasets/Battam/agentic-env-index/revision/main | python3 -c "import json, sys; print(json.load(sys.stdin)['sha'])"
Put it in place of <commit> in the examples below.
- Browse: the dataset viewer has one subset per source.
- Query with DuckDB, for example the ten largest PDF files across all sources:
SELECT source, uri, size_bytes, lower(hex(sha256)) AS sha256 FROM 'hf://datasets/Battam/agentic-env-index@<commit>/data/*/*/*.parquet'
WHERE ext = 'pdf' ORDER BY size_bytes DESC LIMIT 10;
Single files at the same commit: https://huggingface.co/datasets/Battam/agentic-env-index/resolve/<commit>/meta/sources.json.
- Download and verify:
python tools/fetch.py --revision <commit> --help(filters by tag, type, source, license tier and size; checks every hash). See docs/selecting.md.
Sources
| Source | Rows | Size | Top types | Training | Redistribution | Unit | Link check |
|---|---|---|---|---|---|---|---|
Ansible Galaxy collections (ansible-galaxy-collections) |
52,784 | 0 B (52,784 unsized) | gz | yes 70%, unknown 28%, no 2%, conditional 0% | yes 36%, conditional 35%, unknown 29% | file | not checked |
Apache Jira issues (apache-jira) |
1,179,062 | 1.28 GB | txt | unknown | unknown | API response | not checked |
Arquivo.pt office documents (arquivo-pt) |
85,980 | 3.45 GB (82,217 unsized) | pdf, doc, ppt | unknown | no | file | not checked |
Artifact Hub Helm charts (artifact-hub-helm-charts) |
127,281 | 0 B (127,281 unsized) | tgz | unknown 82%, yes 18% | unknown 82%, yes 16%, conditional 2% | file | not checked |
arXiv source packages (arxiv-source) |
3,165,721 | 14.1 TB | gz, pdf, tar | unknown 75%, yes 20%, conditional 5% | no 60%, yes 19%, unknown 14%, conditional 6% | archive member | not checked |
Bioinformatics public data archives (bioinformatics-public-archives) |
278,202 | 7.81 TB (237,457 unsized) | gz, (none), zip | yes 96%, unknown 4% | yes 96%, unknown 4%, conditional 0% | file | not checked |
CentOS Community Build System build logs (centos-cbs-build-logs) |
113,676 | 155 GB | log | unknown | unknown | file | not checked |
Chromium LUCI failed build logs (chromium-luci-builds) |
8,665,027 | 0 B (8,665,027 unsized) | log | unknown | unknown | file | not checked |
Common Crawl office files (common-crawl) |
187,192,070 | 149 TB | pdf, txt, xml | unknown | unknown | WARC record | not checked |
conda-forge packages (conda-forge-packages) |
3,369,245 | 25.4 TB | conda, bz2 | yes 94%, unknown 6% | yes 73%, conditional 22%, unknown 6% | file | not checked |
CPAN distributions (cpan-distributions) |
165,740 | 38.4 GB | gz, tgz, zip | yes 71%, unknown 29% | conditional 64%, unknown 29%, yes 7% | file | not checked |
CRAN packages (cran) |
25,299 | 18.9 GB | gz | yes | conditional 67%, yes 33% | file | 2026-10-02: 100% of 25,284 reachable |
crates.io packages (crates-io) |
3,252,145 | 663 GB | crate | yes 96%, unknown 3%, conditional 1% | yes 87%, conditional 9%, unknown 3% | file | not checked |
CTFtime events, tasks and writeups (ctftime) |
75,162 | 489 MB (33,023 unsized) | html, json | no | unknown | file | not checked |
DataCite DOI metadata (datacite-dois) |
42,611,788 | 2.23 TB (42,609,608 unsized) | json, (none), zip | yes 99%, unknown 1%, conditional 0% | yes 99%, unknown 1%, conditional 0% | API response | not checked |
Debian bug reports (debian-bugs) |
2,244,932 | 145 GB | log, summary | unknown | unknown | file | 2026-10-01: 97% of 60 reachable |
Debian source packages (debian-sources) |
115,614 | 99.2 GB | xz, dsc, gz | yes | conditional | file | not checked |
Debian and Ubuntu binary packages (debian-ubuntu-binary-packages) |
676,230 | 2.11 TB | deb, xz | yes 93%, unknown 7% | conditional 93%, unknown 7% | file | not checked |
Discourse forum posts (discourse-forums) |
13,543,650 | 9.71 GB | md | unknown 57%, conditional 35%, yes 8% | unknown 57%, conditional 40%, yes 2% | API response | not checked |
Docker Hub container images (docker-hub) |
593,368 | 221 TB | json | conditional | unknown | file | not checked |
e-Gov Data Portal (Japan) (egov-japan-data) |
292,050 | 15.4 GB (244,867 unsized) | html, do, xls | yes 99%, unknown 1%, conditional 0% | yes 99%, unknown 1%, conditional 0% | file | not checked |
End of Term Web Archive (end-of-term-web-archive) |
337,638,097 | 149 TB | xml, pdf, txt | yes | yes | WARC record | 2026-10-02: 100% of 180 reachable |
Enron spreadsheets and emails (enron-spreadsheets-emails) |
14 | 9.85 GB | 7z | yes | yes | file | 2026-10-01: 100% of 14 reachable |
EU Publications Office CELLAR documents (eu-cellar-documents) |
1,921,815 | 803 GB (1,396,072 unsized) | pdf, xml, html | yes | yes | file | not checked |
European Parliament plenary records (european-parliament-plenary) |
1,233,677 | 38.0 GB (712,457 unsized) | pdf, xml, doc | yes | yes | file | not checked |
F-Droid apps (f-droid-apps) |
196,376 | 7.30 TB | apk, gz, zip | yes 100%, unknown 0% | conditional 71%, yes 28%, unknown 0% | file | not checked |
fatcat scholarly files (fatcat-files) |
55,929,113 | 80.3 TB (27,845 unsized) | pdf, (none), xml | unknown | unknown | file | not checked |
Fedora Koji build logs (fedora-koji-build-logs) |
193,588 | 257 GB | log | unknown | unknown | file | not checked |
FreeBSD Bugzilla bugs (freebsd-bugzilla) |
573,793 | 5.67 GB | json, diff, patch | unknown | unknown | API response | not checked |
FUSE spreadsheets (fuse-spreadsheets) |
1 | 9.39 GB | zip | yes | yes | file | 2026-10-01: 100% of 1 reachable |
FUSE spreadsheets (Common Crawl records) (fuse-spreadsheets-common-crawl) |
249,376 | 7.59 GB | xls, xlsx, xltx | unknown | unknown | WARC record | not checked |
GCC Bugzilla bugs (gcc-bugzilla) |
186,889 | 4.93 GB (53 unsized) | json, patch, c | unknown | unknown | API response | not checked |
GH Archive hourly event files (gh-archive-hourly) |
136,799 | 5.02 TB | gz | unknown | unknown | file | not checked |
GitHub repositories (selected languages) (github-gap-language-repos) |
8,713,457 | 825 GB | scala, v, (none) | yes 56%, unknown 44% | yes 45%, unknown 44%, conditional 12% | file | not checked |
GitHub release assets (github-release-assets) |
184,200 | 4.27 TB (43,236 unsized) | gz, png, zip | unknown 56%, yes 44% | unknown 56%, yes 35%, conditional 10% | file | not checked |
Go module mirror zips (go-module-proxy-zips) |
59,120,216 | 0 B (59,120,216 unsized) | zip | yes 59%, unknown 41%, conditional 0% | yes 49%, unknown 41%, conditional 9% | file | not checked |
GovDocs1 documents (govdocs1) |
1,000 | 333 GB | zip | yes | yes | file | 2026-10-01: 100% of 1,000 reachable |
GovDocs1 files (govdocs1-files) |
986,279 | 501 GB | pdf, html, jpg | yes | yes | archive member | not checked |
govinfo bulk data (govinfo-bulk-data) |
419,274 | 56.4 GB | xml, zip, html | yes | yes | file | 2026-10-01: 9% of 351 reachable |
Hackage packages (hackage-packages) |
154,127 | 20.0 GB | gz | yes 93%, unknown 7% | yes 84%, conditional 9%, unknown 7% | file | not checked |
Hacker News discussion threads (hacker-news-threads) |
4,790,481 | 0 B (4,790,481 unsized) | json | no | no | file | not checked |
Harvard Dataverse (harvard-dataverse) |
1,104,178 | 68.2 TB | jpg, nc, tif | yes 82%, conditional 18% | yes 80%, conditional 20% | file | not checked |
Homebrew packages (homebrew-packages) |
81,275 | 572 GB (31,254 unsized) | gz, rb, dmg | yes 87%, unknown 13% | yes 58%, conditional 29%, unknown 13% | file | not checked |
Mailman 3 HyperKitty list archives (hyperkitty-mailing-list-archives) |
80,365 | 0 B (80,365 unsized) | gz | unknown | unknown | file | not checked |
IF Archive text adventures (if-archive) |
23,819 | 53.3 GB | zip, png, txt | no 99%, unknown 1% | no 99%, unknown 1% | file | not checked |
Internet Archive software and open media (internet-archive-open-licensed) |
4,131,497 | 219 TB | sig, zst, xz | unknown 98%, yes 2%, conditional 1% | unknown 98%, yes 2%, conditional 1% | file | not checked |
Kaggle datasets and competitions (kaggle-file-listings) |
205,735 | 118 GB | webp, png, tfrec | yes 98%, unknown 2%, conditional 0% | yes 98%, unknown 2%, conditional 0%, no 0% | file | not checked |
KDE, WebKit and Gentoo Bugzilla bugs (kde-webkit-gentoo-bugzilla) |
1,923,669 | 83.0 GB | json, log, txt | unknown | unknown | API response | not checked |
Linux kernel Bugzilla bugs (kernel-bugzilla) |
125,846 | 10.6 GB | json, txt, (none) | unknown | unknown | API response | not checked |
Launchpad Ubuntu bugs (launchpad-ubuntu-bugs) |
97,333 | 76.2 MB (30,111 unsized) | json, (none), log | unknown | unknown | API response | not checked |
Lean toolchains and packages (lean-toolchains-and-packages) |
37,933 | 7.41 TB (872 unsized) | gz, zip, zst | yes 56%, unknown 44% | yes 56%, unknown 44%, conditional 0% | file | not checked |
LibreOffice Bugzilla bugs (libreoffice-bugzilla) |
244,962 | 51.1 GB | json, png, odt | yes | conditional | API response | not checked |
Lichess open database (lichess) |
1,369 | 2.63 TB | zst | yes | yes 94%, conditional 6% | file | 2026-10-01: 100% of 1,369 reachable |
Linux kernel mailing list archives (linux-kernel-mail-archives) |
405 | 0 B (405 unsized) | git | no | unknown | file | 2026-10-01: 0% of 405 reachable |
Loghub system logs (loghub) |
19 | 6.00 GB | gz, zip | conditional | conditional | file | 2026-10-01: 100% of 19 reachable |
Mastodon public posts (mastodon-public-timelines) |
4,914,927 | 1.73 GB (1,144,744 unsized) | html, png, jpg | unknown 61%, no 39% | unknown | API response | not checked |
Mendeley Data research data files (mendeley-data) |
173,005 | 2.65 TB | tif, jpg, txt | yes 93%, conditional 6%, unknown 0% | yes 93%, conditional 7%, no 1%, unknown 0% | file | not checked |
ModelScope datasets and models (modelscope) |
8,726,153 | 2.03 PB | mp4, jpg, parquet | yes 85%, unknown 10%, conditional 5% | yes 83%, unknown 10%, conditional 7%, no 0% | file | not checked |
Mozilla Bugzilla bugs (mozilla-bugzilla) |
3,896,234 | 266 GB | json, txt, patch | unknown | unknown | file | not checked |
Mozilla Treeherder failed job logs (mozilla-treeherder-ci-logs) |
37,531 | 80.4 GB (8,609 unsized) | log | unknown | unknown | file | not checked |
NapierOne and Digital Corpora scenarios (napierone-digital-corpora) |
1,056 | 2.49 TB | zip, e01, xml | yes 97%, unknown 3% | conditional 97%, unknown 3% | file | 2026-10-01: 99% of 307 reachable |
NapierOne files (napierone-files) |
659,576 | 922 GB | webp, jpg, 7z | yes 93%, unknown 7% | conditional 93%, unknown 7% | archive member | not checked |
npm packages (npm) |
55,908,442 | 0 B (55,908,442 unsized) | tgz | yes 81%, unknown 19%, conditional 0% | yes 76%, unknown 19%, conditional 6% | file | not checked |
NuGet packages (nuget) |
807,199 | 625 GB | nupkg | unknown 100%, yes 0% | unknown 100%, yes 0% | file | not checked |
NVD CVE vulnerability records (nvd-cve) |
381,690 | 1.95 GB | json | yes | yes | API response | not checked |
OpenDev Zuul CI job logs (opendev-zuul-ci) |
27,332 | 10.6 GB | gz | unknown | unknown | file | not checked |
OpenStreetMap planet files (openstreetmap-planet) |
27,295,635 | 9.13 TB (14,266,744 unsized) | gz, txt, torrent | yes 100%, unknown 0% | conditional 100%, unknown 0% | file | not checked |
OSV open source vulnerability database (osv-vulnerabilities) |
2,616,111 | 22.5 GB | json, csv, zip | unknown 82%, yes 18% | unknown 82%, yes 15%, conditional 3% | file | not checked |
PubMed Central Open Access articles (pmc-open-access) |
6,135,573 | 1.83 TB | jpg, xml, txt | yes 63%, conditional 29%, unknown 8% | yes 63%, conditional 20%, no 9%, unknown 8% | file | not checked |
Project Gutenberg ebooks (project-gutenberg) |
2,024,683 | 716 GB (942,233 unsized) | png, epub, mobi | yes 99%, unknown 1% | conditional 99%, no 1% | file | not checked |
PyPI packages (pypi) |
45,486,331 | 129 TB | whl, gz, egg | yes 82%, unknown 17%, conditional 0% | yes 70%, unknown 17%, conditional 12% | file | not checked |
Reddit threads (Arctic Shift archive) (reddit-threads-arctic-shift) |
318,856 | 0 B (318,856 unsized) | json | no | no | file | not checked |
Red Hat Bugzilla bugs (redhat-bugzilla) |
1,699,387 | 142 GB | json, (none), log | unknown | unknown | API response | not checked |
Rosetta Code (GitHub community mirror) (rosetta-code-github-mirror) |
148,991 | 153 MB | basic, j, py | unknown | unknown | file | not checked |
RubyGems packages (rubygems) |
552,444 | 451 GB | gem | yes 70%, unknown 30%, conditional 0% | yes 66%, unknown 30%, conditional 4% | file | not checked |
SafeDocs CC-MAIN-2021-31 PDF and UNSAFE (safedocs-cc-main-2021-31) |
13,457,931 | 14.9 TB | pdf, (none), pcap | unknown | unknown | byte range | not checked |
SEC EDGAR company filings (sec-edgar-filings) |
18,788,783 | 0 B (18,788,783 unsized) | txt | yes | yes | file | not checked |
Smithsonian Open Access (smithsonian-open-access) |
9,311,739 | 851 TB | jpg, tif, json | yes | yes | file | not checked |
Solver benchmark instances (solver-benchmark-instances) |
186,738 | 146 GB | xz, c, i | unknown 57%, yes 43% | unknown 57%, yes 43%, conditional 0% | file | not checked |
Sourceware mailing list archives (sourceware-public-inbox) |
168 | 0 B (168 unsized) | git | unknown | unknown | file | not checked |
Stack Exchange recent posts (API) (stackexchange-api-posts) |
3,084,502 | 5.03 GB | html | no | conditional | API response | not checked |
Stack Exchange data dump (stackexchange-dump) |
3,629 | 960 GB | 7z | no 90%, yes 10% | no 90%, conditional 10% | file | 2026-10-01: 100% of 3,629 reachable |
Ubuntu IRC logs (ubuntu-irc-logs) |
456,106 | 51.6 MB (229,171 unsized) | txt | conditional | conditional | file | 2026-10-02: 100% of 8,846 reachable |
UK Parliament Hansard debates (uk-parliament-hansard) |
1,370,783 | 0 B (1,370,783 unsized) | json | yes | yes | API response | not checked |
Wikimedia talk page dumps (wikimedia-talk-dumps) |
1,563 | 503 GB | bz2 | yes | conditional 97%, yes 3% | file | not checked |
World Bank open data files (worldbank-open-data) |
33,928 | 334 GB (484 unsized) | zip, xlsx, csv | yes 99%, conditional 1%, unknown 0% | yes 98%, conditional 1%, unknown 0% | file | not checked |
Zenodo10K slides (zenodo10k-slides) |
10,448 | 242 GB | pptx | yes 98%, conditional 1%, unknown 0% | yes 96%, conditional 4%, unknown 0% | file | 2026-10-01: 100% of 10,448 reachable |
Zulip public chat archives (zulip-public-chats) |
1,300,638 | 239 MB | txt | no | conditional | API response | not checked |
The training and redistribution tiers are the source's license tiers; rows can carry their own tier in the license columns. A row is one file unless the unit says otherwise.
A batch still being collected also has a part-tail.parquet in its folder: it holds the rows that do not yet fill a numbered part file and is replaced as collection goes on (it may be empty for a while); once enough rows arrive they move into the next part-<number>.parquet, so a row's file and row number in the tail are not stable. A finished batch has no tail file: all of its rows are in numbered part files.
Rows are compared with public benchmark tasks in scans that run after collection, so not every row has been compared yet. meta/bench_checked.parquet says which data files were compared and docs/selecting.md says against which benchmarks, up to when, and how the default filter treats rows that were not compared.
More: docs/selecting.md (fields, the default filter and examples); meta/sources.json (per-source details); meta/coverage.json (totals by license tier, source, unit and size); cards/ (one card per source: what its files are, from sample files that were downloaded and read; fields that were judged carry their evidence, fields not judged yet are marked unfilled).
Rows that need a login
Some rows can only be downloaded after a login, a form or an access request. The full pointer says so in access.needs_auth. The value false means no login is needed. The value unknown means the source does not say. Other values name what is needed, for example guestbook (a form), restricted (ask the authors), embargoed (wait until access.embargo_until), requester_pays (your own AWS account pays) or kaggle_token (a Kaggle account).
Most sources set one value for all rows in layout.constants.access of meta/sources.json. A row with another value keeps it in the extra column. This query counts the rows of each source by value:
WITH d AS (
SELECT json_extract_string(s, '$.source') AS source,
json_extract_string(s, '$.layout.constants.access.needs_auth') AS dflt
FROM (SELECT unnest(json_extract(content, '$.sources[*]')) AS s
FROM read_text('hf://datasets/Battam/agentic-env-index@<commit>/meta/sources.json'))),
r AS (
SELECT p.source, p.uri, coalesce(CASE
WHEN json_type(p.extra, '$._row') IS NOT NULL THEN json_extract_string(p.extra, '$._row.access.needs_auth')
WHEN json_type(p.extra, '$.access') IS NOT NULL THEN json_extract_string(p.extra, '$.access.needs_auth')
WHEN json_contains(json_extract(p.extra, '$._absent'), '"access"') THEN NULL
ELSE d.dflt END, 'false') AS needs_auth
FROM 'hf://datasets/Battam/agentic-env-index@<commit>/data/*/*/*.parquet' p LEFT JOIN d USING (source))
SELECT source, needs_auth, count(*) AS rows FROM r GROUP BY ALL ORDER BY ALL;
To keep only rows that need no login, replace the last line with SELECT * FROM r WHERE needs_auth = 'false';. To read one source, replace the first * after data/ with its name.
tools/fetch.py sends no request for these rows. It writes them to its manifest as skipped, with the reason. It does try rows marked unknown. At the end it prints how many rows it skipped for each value. Requester-pays rows are fetched when your AWS keys are set. Embargoed rows are fetched from the date in access.embargo_until.
License
This catalog itself (addresses, sizes, hashes and descriptive text) is released under CC0 1.0; that dedication does not extend to the files it points to. Each file keeps the license of its source.
The columns license_spdx, license_training_ok, license_redistribution_ok and license_notes hold a row's own license. They are filled only when the row has a license of its own. An empty column does not mean the row has no license. It means the row takes the license of its batch. The batch license is batches[].license in meta/sources.json. meta/batches.parquet has the same tiers in its columns training_ok and redistribution_ok. A source can also set one license for all its rows, in layout.constants.license_override of meta/sources.json. That license comes before the batch license. A tier other than yes, conditional or no counts as unknown. tools/fetch.py --tier applies these rules for you.
In these sources license_training_ok and license_redistribution_ok are empty in every row, so every row takes the tiers of its batch license: apache-jira, centos-cbs-build-logs, chromium-luci-builds, common-crawl, ctftime, debian-bugs, debian-sources, docker-hub, end-of-term-web-archive, enron-spreadsheets-emails, eu-cellar-documents, european-parliament-plenary, fedora-koji-build-logs, freebsd-bugzilla, fuse-spreadsheets, fuse-spreadsheets-common-crawl, gcc-bugzilla, gh-archive-hourly, govdocs1, govdocs1-files, govinfo-bulk-data, hacker-news-threads, hyperkitty-mailing-list-archives, kde-webkit-gentoo-bugzilla, kernel-bugzilla, launchpad-ubuntu-bugs, libreoffice-bugzilla, linux-kernel-mail-archives, loghub, mozilla-bugzilla, mozilla-treeherder-ci-logs, nvd-cve, opendev-zuul-ci, reddit-threads-arctic-shift, redhat-bugzilla, rosetta-code-github-mirror, sec-edgar-filings, sourceware-public-inbox, stackexchange-api-posts, stackexchange-dump, ubuntu-irc-logs, uk-parliament-hansard, zulip-public-chats.
The catalog itself contains no personal information. Some of the files it points to, such as email archives and Q&A posts, contain names and contact details that their authors published; follow each source's terms when you use them.
- Downloads last month
- 2,327