# Source card (card_format training-data-foundation). Readers: agents that choose sources and fetch files to build # terminal tasks. Field names follow the corpus card format where the meaning is the same; provenance is where # the data comes from and its licence, labeling is who judged the card. # evidence..fill: auto (computed from the index), judgment (from sample files downloaded and read), # derived (fixed rule over other fields), unfilled (no sample evidence yet, left empty on purpose; basis says # why and the plan), not_applicable (this class of card never has a value for the field; basis says why). # Sample numbers #NN refer to sample.files. Ratios are counts in the sample, not estimates for the source. card_version: 5 card_format: training-data-foundation source: modelscope parent: null title: null description: null catalog: agentic-env-index snapshot: rev: "6d33b190f91743767248c7a808c5fa8bb8954206" at: "2026-10-08" rev_at: "2026-10-04T07:16:55Z" header_note: "Still being collected: no collection header yet, so title, description, licence and fetch fields are empty; counts are the pointer shards present at the snapshot date and will grow." audit: {status: draft, reviewed_by: null} labeling: {labeler: null, model: null, brief_sha: null} status: flagged status_reasons: ["no collection header yet: title, description, licence and fetch fields are empty"] duplicate_of: null prior_taxon: null provenance: url: null licence: null licence_url: null per_collection: null licence_tier_counts: {permissive: 7312659, share-alike: 139536, nc-research: 456990, unknown: 902619} training_ok: mixed redistribution_ok: unknown attribution: null per_file: null licence_tiers: yes/yes: 7312659 unknown/unknown: 902185 conditional/conditional: 452062 yes/conditional: 139536 conditional/no: 4928 unknown/conditional: 434 spdx_top: Apache-2.0: 5996814 "(none)": 948502 CC-BY-4.0: 588502 MIT: 428302 CC-BY-NC-4.0: 223357 ODC-By-1.0: 204905 CC-BY-NC-SA-4.0: 178157 CC-BY-SA-4.0: 65774 GPL-3.0: 39172 CC0-1.0: 27696 CC-BY-NC-ND-4.0: 17113 CC-BY-3.0: 15477 AGPL-3.0: 14957 BSD-3-Clause: 12320 CC-BY-NC-2.0: 10078 spdx_distinct: 29 licence_sample_check: null license: {note: null, tier: mixed, ship: mixed} unit: pointer_level: null units: {file: 8811804} state: null expanded_count: null rows_without_size: 0 collections: distinct: 6018 top: "dataset/Ran0501/JieZi@dd6b3dd7": 10000 "dataset/RoboDojo-Benchmark/GOAI-2026@4a7f09b5": 10000 "dataset/X-Humanoid/RoboMIND2.0-Franka-sim@412fa253": 10000 "dataset/lerobot/abc_130k_v3_train@5b1e1b0e": 10000 "dataset/NICDFcau/NICDF-AquaDetSeg@11c03d14": 10000 "dataset/whynlp/SlimPajama-627B-Llama-tokenized@6293a176": 10000 "dataset/WallelWan/OMA@04508779": 10000 "dataset/L67plus/TASLE@8184df8e": 10000 "dataset/Xinary/KONIQ@2fdc68bc": 10000 "dataset/yifengzhan/OVCOD-D@649813da": 10000 sub_collections: null batches: null n_ext: 7531 usable_unit: null size: max: 206262766768 p50: 487756 p95: 897228995 bytes: 2074641965322165 files: 8811804 truncated_listing: false form: by_ext: mp4: 1435485 jpg: 1379150 parquet: 1215538 png: 937992 json: 509719 wav: 347786 hdf5: 272747 tar: 260694 txt: 232521 zip: 187139 gz: 168026 npy: 136603 jsonl: 115434 zst: 83611 lock: 81096 metadata: 78500 npz: 74527 md: 68773 pdf: 65908 mp3: 65548 "(7511 other types)": 1095007 size_quantiles: p25: 43793 p50: 487756 p75: 12401490 p90: 335403406 p95: 897228995 p99: 4393873309 p99.9: 23327477422 min: 0 max: 206262766768 method: "nearest rank: sort sizes ascending, take the k-th with k = max(1, ceil(q * n)); rows without a size are left out" by_ext_size: - {ext: mp4, rows: 1435485, bytes: 45138859600605, p50: 4224519, p95: 67512971, max: 28959866383} - {ext: jpg, rows: 1379150, bytes: 651766061547, p50: 95087, p95: 1598099, max: 171096186} - ext: parquet rows: 1215538 bytes: 473081641183208 p50: 44372280 p95: 2153695335 max: 56402186977 - {ext: png, rows: 937992, bytes: 976051231609, p50: 240098, p95: 2921782, max: 7222120979} - {ext: json, rows: 509719, bytes: 1306819889886, p50: 5132, p95: 825692, max: 38810371206} - {ext: wav, rows: 347786, bytes: 813619467065, p50: 330882, p95: 12017768, max: 794977150} - {ext: hdf5, rows: 272747, bytes: 86245931918467, p50: 142002073, p95: 979880599, max: 91127786704} - ext: tar rows: 260694 bytes: 634445541080445 p50: 1045544960 p95: 8429926400 max: 107366062080 - {ext: txt, rows: 232521, bytes: 86326683771, p50: 234, p95: 309593, max: 17363908882} - {ext: zip, rows: 187139, bytes: 174869559099593, p50: 29558168, p95: 5122802134, max: 111643559279} - {ext: gz, rows: 168026, bytes: 122450677014603, p50: 8501194, p95: 3115120769, max: 138666415688} - {ext: npy, rows: 136603, bytes: 2436888745794, p50: 5243008, p95: 44826752, max: 53033327335} - {ext: jsonl, rows: 115434, bytes: 61209061333722, p50: 20345901, p95: 1783013504, max: 52243347728} - {ext: zst, rows: 83611, bytes: 37628610547754, p50: 77971582, p95: 3539649139, max: 142985307495} - {ext: lock, rows: 81096, bytes: 385833, p50: 0, p95: 0, max: 320394} - {ext: metadata, rows: 78500, bytes: 9698177, p50: 125, p95: 125, max: 125} - {ext: npz, rows: 74527, bytes: 3012828716585, p50: 612156, p95: 114494525, max: 9336597661} - {ext: md, rows: 68773, bytes: 780035483, p50: 2588, p95: 45085, max: 1549661} - {ext: pdf, rows: 65908, bytes: 599936852379, p50: 3881039, p95: 36400596, max: 1090574422} - {ext: mp3, rows: 65548, bytes: 89228047416, p50: 219405, p95: 7402605, max: 104076060} over_cap: cap: 61944279 rows_over: 1555185 bytes_over: 2047654402670199 note: "Default fetch-time cap (the overall p99 of benchmark task files, provisional). It only stops a download; it removes no rows from the index or the default subset." largest: - pointer_id: "modelscope:dataset/arcinstitute/State-Tahoe-Filtered@5bd767d9:c22.h5ad" size_bytes: 206262766768 - pointer_id: "modelscope:dataset/arcinstitute/State-Tahoe-Filtered@5bd767d9:c19.h5ad" size_bytes: 183498201272 - pointer_id: "modelscope:dataset/arcinstitute/State-Tahoe-Filtered@5bd767d9:c20.h5ad" size_bytes: 157840480424 - pointer_id: "modelscope:dataset/IvanHU/dolma3-2T@297cc4cf:worker_112_input_ids_tmp.bin.zst" size_bytes: 142985307495 - pointer_id: "modelscope:dataset/JEdward/AHA@da430c97:stack_blocks.tar.gz" size_bytes: 138666415688 date_years: null media_type: {"(none)": 8811804} facets: null artifact: null source_origin: null parse_cost: null parse_quirks: null structure: self_describing: null integrity_builtin: null has_reference_output: null invariants: null tools: null ships: problem_statement: {present: null} reference_solution: {share: null, present: null} tests: {present: null} env_definition: {present: null} known_answer: {present: null} generator: {present: null} upstream_labels: {present: null} class: null subject: domain_span: null shape_span: null skill_family_span: null ont_domain: null tech_vocabulary: null label: summary: null source_kind: null notes: null excluded_parts: null domain_span: null contamination: scope: "25590 of 8811804 pointers removed (26217 flagged)" benchmark: "APEX-Agents, DeepSWE, FrontierSWE v2, GDPval, OSWorld, OSWorld 2, Terminal-Bench, agentslastexam, swebench, tau2-bench" evidence: "pointer names, paths and contents matched against the task files and data addresses of public agent benchmarks; benchmark copies, task files and data addresses are removed, and so are whole containers that copy a benchmark or are derived from one" unit_kinds: null derived: {coverage: null, unit_kinds: null} risk: contamination: tb_overlap overlap_check: done: true pointers_flagged: 26217 pointers_removed: 25590 by_class: content_copy_container: 11961 derivative_container: 13576 generic_text: 190 task_file: 53 trivial_small: 437 benches: APEX-Agents: 4726 DeepSWE: 800 "FrontierSWE v2": 471 GDPval: 552 OSWorld: 730 "OSWorld 2": 595 Terminal-Bench: 3592 agentslastexam: 12105 swebench: 1297 tau2-bench: 900 method: "pointer names, paths and contents matched against the task files and data addresses of public agent benchmarks; benchmark copies, task files and data addresses are removed, and so are whole containers that copy a benchmark or are derived from one" pii: null pii_evidence: null malware: null ship_bytes: unknown subset: rule: "take all rows; licence, size and version are labels, not filters" rows_in_subset: 8811804 last_built: null access: method: null verify: null hosts: {"https://modelscope.cn": 8811804} uri_example: "https://modelscope.cn/api/v1/datasets/0xSero/structured-outputs-calibration-v1/repo?Revision=2b93dd5b&FilePath=.gitattributes" needs_auth: {"false": 8811804} credential: null hash_source: {source_published: 8811804} range_supported: null reachable: index_flag: {unknown: 8811804} index_check: null our_sample: null drift: null sample: null probe: null checks: errors: [] warnings: ["no collection header yet: title, description, licence and fetch fields are empty"] history: [] evidence: access.credential: {fill: unfilled, basis: "no judgment file for this source yet"} access.drift: {fill: unfilled, basis: "no judgment file for this source yet"} access.range_supported: {fill: unfilled, basis: "no judgment file for this source yet"} card_format: fill: derived basis: "fixed name of this card format; the corpus cards are another format and keep their own card_version numbers" checks.errors: fill: derived basis: "a card that fails a check (judgment file validation, public text check) is not written, so a written card has none" checks.history: fill: not_applicable basis: "the card generator has no repair rounds: a failing check stops the build instead of being repaired, so there is never a round to record; this holds only while the card generators have no repair step: if one is added, this entry leaves the list" checks.warnings: fill: derived basis: "data problems found while building the card (missing header, unreadable or absent shards, no overlap scan)" class: {fill: unfilled, basis: "ships not judged"} derived.coverage: fill: unfilled basis: "computed from label.unit_kinds and the pointer rows; label.unit_kinds is not judged yet; plan: left to a later projection step (derived.* is computed there, not on the card)" derived.unit_kinds: fill: unfilled basis: "computed from label.unit_kinds and the pointer rows; label.unit_kinds is not judged yet; plan: left to a later projection step (derived.* is computed there, not on the card)" duplicate_of: fill: derived basis: "the catalog keeps one card per source name and writes no second card for a source, so there is no card to point to" form.artifact: {fill: unfilled, basis: "no judgment file for this source yet"} form.by_ext: fill: auto basis: "computed from every pointer row of this source (licence tiers after the current per-row licence rules)" checked_at: "2026-10-08" form.date_years: fill: auto basis: "computed from every pointer row of this source (licence tiers after the current per-row licence rules)" checked_at: "2026-10-08" form.over_cap: fill: auto basis: "computed from every pointer row of this source (licence tiers after the current per-row licence rules)" checked_at: "2026-10-08" form.parse_cost: {fill: unfilled, basis: "no judgment file for this source yet"} form.parse_quirks: {fill: unfilled, basis: "no judgment file for this source yet"} form.size_quantiles: fill: auto basis: "computed from every pointer row of this source (licence tiers after the current per-row licence rules)" checked_at: "2026-10-08" form.source_origin: {fill: unfilled, basis: "no judgment file for this source yet"} label.contamination.benchmark: {fill: derived, basis: "from risk.overlap_check (benchmark overlap scan over all pointer rows)"} label.contamination.evidence: {fill: derived, basis: "from risk.overlap_check (benchmark overlap scan over all pointer rows)"} label.contamination.scope: {fill: derived, basis: "from risk.overlap_check (benchmark overlap scan over all pointer rows)"} label.domain_span: fill: unfilled basis: "subject.domain_span is not judged yet; plan: judged from a stratified sample by one model, with key fields re-judged by a second model" label.excluded_parts: fill: unfilled basis: "added in card_version 2 and not judged from the sample yet; plan: judged from a stratified sample by one model, with key fields re-judged by a second model" label.notes: fill: unfilled basis: "added in card_version 2 and not judged from the sample yet; plan: judged from a stratified sample by one model, with key fields re-judged by a second model" label.source_kind: fill: unfilled basis: "added in card_version 2 and not judged from the sample yet; plan: judged from a stratified sample by one model, with key fields re-judged by a second model" label.summary: fill: unfilled basis: "added in card_version 2 and not judged from the sample yet; plan: judged from a stratified sample by one model, with key fields re-judged by a second model" label.unit_kinds: fill: unfilled basis: "added in card_version 2 and not judged from the sample yet; plan: judged from a stratified sample by one model, with key fields re-judged by a second model" labeling.brief_sha: fill: unfilled basis: "not recorded; plan: labeling is written when a judged file records who labelled it" labeling.labeler: fill: unfilled basis: "not recorded; plan: labeling is written when a judged file records who labelled it" labeling.model: fill: unfilled basis: "not recorded; plan: labeling is written when a judged file records who labelled it" license.ship: fill: derived basis: "from license.tier by the one fixed table (permissive yes/yes, share-alike conditional/yes, nc-research no/conditional, internal-only no/unknown, unknown no/unknown, no-training no/no; mixed when the tier is mixed); ship and training_ok are listed as (ship, training_ok)" license.tier: fill: derived basis: "the licence is set per row (or differs between collections), so the header says mixed; provenance.licence_tier_counts gives the rows per tier; tier words are the corpus card's, plus our own no-training tier" prior_taxon: fill: not_applicable basis: "the prior taxon is a code from the seed list the corpus cards were started from; sources in this catalog were not drawn from that list and never are: no step that adds sources draws from it; that list is private material, so even if the same source is also in it one day, its code does not go on a public card" probe: fill: unfilled basis: "no review result recorded for this source yet; plan: filled from our own review (our own review, not a second model re-reading the whole card: two_model counts how many judged fields an independent second model agreed with; spot_check is the conclusion and date of our own spot check)" provenance.licence: {fill: auto, basis: "collection header licence block", checked_at: "2026-10-08"} provenance.licence_sample_check: {fill: unfilled, basis: "no judgment file for this source yet"} provenance.licence_tier_counts: fill: derived basis: "pointer rows per tier, from provenance.licence_tiers by the fixed pair-to-tier rule" provenance.licence_tiers: fill: auto basis: "computed from every pointer row of this source (licence tiers after the current per-row licence rules)" checked_at: "2026-10-08" provenance.spdx_top: fill: auto basis: "computed from every pointer row of this source (licence tiers after the current per-row licence rules)" checked_at: "2026-10-08" provenance.training_ok: fill: derived basis: "from license.tier by the one fixed table (permissive yes/yes, share-alike conditional/yes, nc-research no/conditional, internal-only no/unknown, unknown no/unknown, no-training no/no; mixed when the tier is mixed); ship and training_ok are listed as (ship, training_ok)" risk.contamination: fill: auto basis: "benchmark overlap scan over all pointer rows: tb_overlap when a removed pointer matches a Terminal-Bench task, public_benchmark when any pointer is removed, none otherwise" checked_at: "2026-10-08" risk.malware: {fill: unfilled, basis: "no judgment file for this source yet"} risk.pii: {fill: unfilled, basis: "no judgment file for this source yet"} risk.pii_evidence: {fill: unfilled, basis: "no judgment file for this source yet"} risk.ship_bytes: {fill: derived, basis: "equal to provenance.redistribution_ok"} ships.env_definition.kind: {fill: unfilled, basis: "no judgment file for this source yet"} ships.env_definition.present: {fill: unfilled, basis: "no judgment file for this source yet"} ships.env_definition.runs: {fill: unfilled, basis: "no judgment file for this source yet"} ships.generator.present: {fill: unfilled, basis: "no judgment file for this source yet"} ships.known_answer.present: {fill: unfilled, basis: "no judgment file for this source yet"} ships.problem_statement.present: {fill: unfilled, basis: "no judgment file for this source yet"} ships.reference_solution.present: fill: unfilled basis: "recorded only as {k, n} over label.unit_kinds; the unit kinds of this source are not judged yet; plan: judged from a stratified sample by one model, with key fields re-judged by a second model" ships.reference_solution.share: {fill: unfilled, basis: "no judgment file for this source yet"} ships.tests.coverage: {fill: unfilled, basis: "no judgment file for this source yet"} ships.tests.present: {fill: unfilled, basis: "no judgment file for this source yet"} ships.tests.runs: {fill: unfilled, basis: "no judgment file for this source yet"} ships.upstream_labels.fields: {fill: unfilled, basis: "no judgment file for this source yet"} ships.upstream_labels.present: {fill: unfilled, basis: "no judgment file for this source yet"} size.files: fill: auto basis: "computed from every pointer row of this source (licence tiers after the current per-row licence rules)" checked_at: "2026-10-08" size.truncated_listing: fill: derived basis: "true when some pointer shards were unreadable or absent, so the counts do not cover every row" snapshot.rev: fill: auto basis: "commit of the most recent upload of a data file of this source to the catalog repository, as read back from the repository at upload time; rev_at is its time (UTC)" checked_at: "2026-10-08" status: fill: derived basis: "flagged when checks.warnings is not empty, else ok; review state is audit.status" status_reasons: {fill: derived, basis: "equal to checks.warnings"} structure.has_reference_output: {fill: unfilled, basis: "no judgment file for this source yet"} structure.integrity_builtin: {fill: unfilled, basis: "no judgment file for this source yet"} structure.invariants: {fill: unfilled, basis: "no judgment file for this source yet"} structure.self_describing: {fill: unfilled, basis: "no judgment file for this source yet"} structure.tools: {fill: unfilled, basis: "no judgment file for this source yet"} subject.domain_span: {fill: unfilled, basis: "no judgment file for this source yet"} subject.ont_domain: {fill: unfilled, basis: "no judgment file for this source yet"} subject.shape_span: {fill: unfilled, basis: "no judgment file for this source yet"} subject.skill_family_span: {fill: unfilled, basis: "no judgment file for this source yet"} subject.tech_vocabulary: {fill: unfilled, basis: "no judgment file for this source yet"} subset: fill: auto basis: "default subset rule; last_built is the coverage table of the last subset build" checked_at: "2026-10-08" unit.pointer_level: {fill: unfilled, basis: "no judgment file for this source yet"} unit.state: {fill: unfilled, basis: "pointer_level not judged"} unit.usable_unit: {fill: unfilled, basis: "no judgment file for this source yet"}