training-data-foundation / docs /for-agents.md
Battam's picture
Release [run release-20261111T153845Z]
82d0b24 verified
|
Raw History Blame
12.7 kB

Back to README.md

For agents

English | 简体中文

  • The index holds two layers over the same public web, kept apart and never added together. Corpora are whole upstreams read in bulk and counted in rows; a material set is admitted for building tasks from works the search found upstream or from a group of corpus rows, its files checked and grouped into units, and it is counted in units, apart from any corpus its rows came from. Task material comes first: material sets, grouped into collections, one per task family. A material set (task-ready) is the units admitted from one source (a site, a repository or a dataset) for building tasks (the catalog column kind is material_set, and collection names its collection). A unit is a group of files that its upstream publishes as one thing, delivered whole; its key unit_id is the unit_group its rows share (extra.source_meta.unit_group), defined once, in the glossary of the index's README.md. A unit is material, not a task: one unit may give several tasks, and one task may use several units. Corpora come second, to search when no material set fits: a corpus (raw; plural corpora) is one whole upstream that a collector keeps reading, such as Common Crawl's office documents, PyPI's release files, Mendeley Data or NVD: its rows are pointers to files or API responses, and it is counted in rows (kind is corpus); tools/fetch.py takes a corpus row by row, each under its own key as its unit_id.
  • Pick material sets or corpora from meta/catalog.parquet (one row per material set or corpus, about 50 KB; do not download meta/ whole). Columns: source; kind, material_set or corpus (select one layer: units and rows are never added together); collection, the code of a material set's collection (every code with its name is in meta/vocabulary.json), or of the collection that lists a corpus as holding its kind of material (corpora in meta/vocabulary.json; empty for any other corpus); domain, a list of codes D1 to D10 or leaves such as D3.5, or multi, unclassified; form, such as test-suites or spreadsheets; units (a material set's distinct unit_id values: a test case, one workbook, one package; a corpus is counted in rows, and gives the same number as files) and files (its rows: for a material set, the files of its units and its README.md and LICENSE); publishers (a material set's distinct publisher names over all its rows; empty when its rows name none); language (the language codes of the rows' labels, what the default profile filters on; not recorded when a row has none); contamination.level (by the rule of the default profile of tools/fetch.py over meta/bench_overlap.parquet and meta/bench_checked.parquet: none when every data file is compared and kept and no row is removed, the hit class with the most removed rows when some are removed, not scanned otherwise); license.ship; pii.level; distinct_bytes (deduplicated size) and bytes (recorded size, the sum over rows); size_check (the rows read in meta/sizes.parquet, compared those that compared a size or hash, and how many have each status, match when the measured size agrees; the check's share is match over compared), unsized_rows (rows without a recorded size) and unsized_estimate (their estimated size, with its 95% range); fetch_check (whether the upstream still serves the files: rows, the sampled rows read in meta/sizes.parquet; fetched, those that fetched and matched their recorded size and hash; read_at, the day read; level, all, some, none or not read); the same figures for the whole index, each layer and each collection are in meta/size_figures.json. Who published each file of a material set, and under which license and license URL, is in meta/attribution.parquet (one row per source, publisher, license and license URL, with its number of files). The vocabulary is meta/vocabulary.json: every domain code with its name (and zh, its Chinese name), what it covers and its parent, every form with its Chinese name and what it means, every category of meta/not_included.parquet with its requirement in English and in Chinese, the two layers with their relation, and every collection with its name, what it holds, its campaigns (take a collection with --tag campaign:<name> for each of its campaigns) and the corpora it lists, if any (take one with --source).
  • Size of this release: Whole index: Recorded size 4.45 PB; deduplicated size 4.11 PB; size check: 99.6% match (34,707 of the 35,586 rows read compared); not compared: address unreachable 449, file gone 80, login needed 329, no size from the server 19, other error 2; estimated size of 215,326,513 of the 312,300,338 rows without a recorded size: 202 TB (an estimate; 95% range 171 TB to 245 TB, the sum of its sources' ranges, wider than a joint interval; from 14,161 rows read), the other 96,973,825 with too few rows read for an estimate. Task material: Recorded size 2.88 GB; deduplicated size 2.81 GB; size check: 99.6% match (3,984 of the 3,992 rows read compared); not compared: login needed 7, other error 1. Corpora: Recorded size 4.45 PB; deduplicated size 4.11 PB; size check: 99.6% match (30,723 of the 31,594 rows read compared); not compared: address unreachable 449, file gone 80, login needed 322, no size from the server 19, other error 1; estimated size of 215,326,513 of the 312,300,338 rows without a recorded size: 202 TB (an estimate; 95% range 171 TB to 245 TB, the sum of its sources' ranges, wider than a joint interval; from 14,161 rows read), the other 96,973,825 with too few rows read for an estimate. Distinct content is the sum of size_bytes over the rows minus the sizes of the rows that repeat another row's content (the same sha256, else sha1, else md5 and size, within the same hash scope; one copy of each content is kept, in the first source by registry order), counted after each release over the data files it published, per source in the catalog column distinct_bytes; archive members whose archive the same source also lists whole have other hashes, so they are not subtracted. tools/fetch.py removes duplicates over your own selection when it reads it. The deduplicated sizes here were counted over the data files of release release-2026-11-11.10. The recorded size is the sum of size_bytes over the rows that have one (catalog column bytes). The check reads a sample of rows of every data file (meta/sizes.parquet): 20 of its rows with a recorded size, and of its rows without one its share of its source's sample, sized for their mean size within 25% at 95% confidence by the spread of the sizes read so far (at most 400 rows per source per round, 20 for a source with none read yet); its share is the rows that match their recorded size out of the rows it compared (status match, size_mismatch or hash_mismatch), and the rows it could not compare, which never reached their file, are counted beside it by status (unreachable, gone, needs_auth, no_length, error; catalog column size_check). The estimate for a source's rows without a recorded size (catalog columns unsized_rows and unsized_estimate) is their number times the mean measured size of such rows read, with a 95% BCa bootstrap range (scipy.stats.bootstrap, its default resample count); a source with fewer than 10 such rows read has none. Above one source, the estimate is the sum of its sources' estimates and its range the sum of their ranges, which is wider than a joint interval. The figures of every level are in meta/size_figures.json.
  • Fetch with one command: uv run https://huggingface.co/datasets/Battam/training-data-foundation/resolve/main/tools/fetch.py --domain D3 --form spreadsheets --sample 3 --out ./data. Without uv: pip install duckdb, download tools/fetch.py and run it with python. Select with --source, --kind (material_set or corpus: one layer), --domain, --form, --tag, --ext, or --selection <file> (a selection file the data page saves: the ids of its sources and its commit); --sample N takes N units per source, --limit N takes N units over the whole selection, --unit <unit_id> takes the unit of that unit_id whole (with --source); before its first download the run prints the units, files and total size it selected.
  • It pins the commit it reads and prints it; pass --revision <commit> to get the same files later. A name that does not exist stops with exit code 2 and the closest names.
  • The default profile removes duplicate copies, rows that matched a benchmark (copies of benchmark data, rows of groupings awaiting a ruling), rows not compared with the benchmark list yet (counted apart, with their own reason), files whose license does not allow shipping (license.ship yes and conditional stay), rows recorded in a language other than English, personal data (catalog pii.level present or suspected, rows marked sensitive), and files above 20 MB. Before downloading it prints how many units each criterion removed; each criterion has its own option (--keep-duplicates, --allow-benchmark, --allow-not-compared, --any-license, --any-language, --allow-personal-data, --any-size; --allow-benchmark also keeps rows not compared yet), and --all turns off every criterion except benchmark and not compared, so rows that matched a benchmark or were not compared yet come in only with their own options. Compared with the benchmark list (keys of 2026-11-08): 1,805 of 1,805 material sets; 139 of 139 corpora. The profile also removes every row whose data file was not compared (--allow-benchmark keeps them, unchecked).
  • Output: ./data/<domain>/<source>/<unit folder>/... holds each unit whole, in a folder of its own named after its unit_id, its files at their paths within the unit; ./data/<domain>/<source>/ also holds the README.md and LICENSE of a material set; ./data/_meta/manifest.jsonl has one line per file (source, path, url, sha256, bytes, spdx, licence_url, publisher, retrieved, origin_kind, leaf, unit_id, verified); ./data/_meta/sizes.tsv one line per source. Every file is checked against the sha256 the index records. A unit stored inside a container (such as a disk image inside a zip) is written as the unit itself: fetch.py takes the member out and checks it against the unit's recorded sha256, so no second tool is needed; only units marked expand (archives meant to be used as folders) are unpacked by tools/to_pool.py.
  • The index holds material only: what a task checks is decided when the task is built from the material. Keep out of the environment the agent works in whatever the task checks its result against:
  • Workbooks (such as the form spreadsheets): a published workbook holds the cached results of its formulas, so a workbook put in as fetched hands the agent the values a task may ask it to compute. Strip the cached values first: in .xlsx and .xlsm the <v> of every formula cell (xl/worksheets/*.xml), the chart caches (c:numCache and c:strCache in xl/charts/*.xml), the pivot caches (xl/pivotCache/, definitions and records) and the external link caches (xl/externalLinks/); in .ods the value attributes (office:value and the like) and the text of formula cells; in .xls the cached result of every FORMULA record and its STRING record.
  • Expected results inside a unit (such as test suites, detection rules with their tests, rules with worked examples, and reproductions and evaluation results with the numbers their authors published): the unit carries its own check (expected outputs, expected findings, the worked result, the reported numbers, and the files that hold the same numbers, such as result tables and logs). Keep those files or parts out of the environment and use them only to check the agent's result.
  • The form of a task built from these files (such as a Terminal-Bench style task with environment/, solution/ and tests/):
  • solution/solve.sh is in bash, with the work spread over top-level commands (8 or more), each command that changes the environment on its own line; not wrapped in a function, main, a large heredoc or a single python call.
  • The task runs without network; its data is inside environment/.
  • Written files carry no timestamps, random values or host names; versions are pinned.
  • No background service that has to outlive a step, no database service.
  • Tests check the final state, not only that files exist.