docs/source_collection_matrix.md

Source collection matrix

Status snapshot: 2026-08-12 America/Los_Angeles. The authoritative source list is registry/sources.json (16 entries). Downloaded means that source payload is present under data/cache; API responses, census tables, and Git commit metadata alone are not payload. This inventory does not itself grant permission or constitute legal advice.

Permission values are intentionally conservative. A source-level dataset license does not override embedded repository/file licenses, inherited benchmark material, or source-snippet terms. No record-level provenance decisions have yet been instantiated under schemas/provenance.schema.json, so only explicit prohibitions are final. Conditional means that the stated license appears compatible in principle, but Code-ATLAS must still complete the listed scope, attribution, lineage, and/or legal checks before training or public payload release.

Registered sources

Source	Artifact kind	Intended role	Pinned revision	Downloaded	ML training permission	Public payload redistribution	Current decision
The Stack v3 train (hf_stack_v3_train)	Repository snapshots	Raw pretraining; held-out loss	a34523fd426a429b7686a5e5786496bf2e1b1172	One of 8,192 Parquet shards downloaded for pipeline validation; full payload not collected	Conditional — ODC-By-1.0 at dataset level; every file must pass the project license allowlist; no_license is excluded	Conditional — per-file/source terms and attribution must pass	Public metadata now; payload only for record-level approved files
The Stack v3 full (hf_stack_v3_full)	Raw repository metadata and blobs	Metadata-only; possible legacy-language supplement	Release label 1.0.0; immutable object manifest still pending	No	Unknown / legal review — ODC-By declaration does not resolve file/source terms	Unknown / legal review — custom filtering, provenance, and license audit required	Public metadata only; no payload
Project CodeNet (ibm_project_codenet)	Online-judge submissions	Parallel/task auxiliary; possible functional evaluation	Payload 1.0.0, SHA-256 7e966203…620fb; tooling 55c323b527ab3d6510d55edff188bc0a0d7bc5e5	Yes; full 8.34 GB archive and metadata census	Conditional — the paper's datasheet appendix B.6 confirms CDLA-Permissive-2.0	Conditional — the downloaded tar has no packaged license notice, so a derivative must supply the required agreement text and preserve provenance	Public metadata now; parallel payload only after problem-family split, decontamination, and license packaging
SWE-rebench V2 (nebius_swe_rebench_v2)	Issue/patch/test tasks with repository pointers	Task/SFT candidate; functional eval; benchmark firewall	Dataset 475dd5e8703bb5fb22dd3c60b5d038b019eba1e0; builder c71902a8cf8d2b725f63d51f199f4d3e56f68d2d	Yes; complete pinned HF snapshot	Conditional — task dataset CC-BY-4.0, but row-level repository terms still control derived code/patches	Conditional — attribution plus per-repository license review	Public metadata; payload only after repo allowlist and task-lineage split
SWE-rebench V2 PRs (nebius_swe_rebench_v2_prs)	PR/patch/test tasks with repository pointers	Task/SFT candidate; benchmark firewall	40faf2c1bb160de625f3c3270ac9d62ea45f3f9c	Yes; complete three-shard pinned snapshot and audit	Conditional — CC-BY-4.0 declaration plus unresolved per-repository terms	Conditional — same scope and attribution review	Public metadata; post-training payload not yet approved
SWE-rebench leaderboard (nebius_swe_rebench_leaderboard)	Dynamic executable benchmark	Evaluation; benchmark firewall	34d5a58864acf91613740a09ec5d205228dcfa39	Yes; complete 18-file pinned HF snapshot	No for Code-ATLAS — project policy is never_train regardless of license	Unknown / legal review — CC-BY-4.0 dataset, but per-repository terms apply and cached Parquet lacks the card-described license field	Quarantined evaluation; public metadata/pointers only
SWE-rebench OpenHands trajectories (nebius_swe_rebench_openhands_trajectories)	Agent trajectories	Conditional task/SFT; metadata-only	35455389ab51bf5e2306bfd436ef72d0f98bf882	Yes; complete pinned Parquet snapshot and audit	Conditional — CC-BY-4.0 declaration; derived SWE-task/repository rights and benchmark overlap unresolved	Unknown / legal review	Public metadata; quarantine pending overlap and license firewall
McEval (mceval)	Multilingual executable benchmark	Evaluation	Dataset c4767fec950de7decd487a40a4dab2795df4a094; tooling 09dec129dd5756447227239b3e5b47e543c2e0fd	Yes; complete 541-file pinned HF snapshot	No for Code-ATLAS — evaluation quarantine	Conditional / legal review — card declares CC-BY-SA-4.0; verify artifact scope and satisfy attribution/share-alike before any mirror	Quarantined evaluation; public metadata/pointers preferred
McEval-Instruct (mceval_instruct)	Multilingual instruction-tuning corpus	SFT/task training	57bd46727e49ac3e2890218a146d35c2454a8496	Yes; complete pinned HF snapshot (35,943 rows)	Conditional — CC-BY-SA-4.0 declaration; source-snippet rights and McEval/raw-code lineage unresolved	Conditional / legal review — attribution/share-alike and source scope must pass	Public metadata; quarantine payload until lineage/license audit
mHumanEval (mhuman_eval)	Translated HumanEval benchmark	Evaluation; possible parallel auxiliary	Commit b4ac5b3f59bb8a9b27fabd2d63080963995b1388, SHA-256 f49726a4…5724	Yes; complete exact-commit tar stream audited	No for standard evaluation families; any auxiliary use requires a separate HumanEval-family split and inherited-material review	Unknown / legal review — archive contains Apache-2.0, but inherited HumanEval material remains unresolved	Incomplete-release quarantine: only 17 partial PLs/544,911 rows observed; COBOL and Fortran absent
MultiPL-E (multiple)	Translated HumanEval/MBPP benchmark	Metadata-only; evaluation pointer	3025a531af7450e7df8b96fe0440e9804480bbad	No; commit metadata only	No — repository license explicitly prohibits using its contents as ML training data	Yes only under its modified BSD conditions, but Code-ATLAS will not mirror payload while using it as a benchmark firewall	Quarantined: no training; public metadata/pointer only
xCodeEval (xcodeeval)	Competitive-programming tasks and submissions	Possible parallel/task auxiliary or evaluation; metadata-only	4993b99b6a97e5e19cfa7d4cd75e12545580ad27	No; commit metadata only	Unknown / legal review — CC-BY-NC-4.0 excludes commercial use and source terms remain unresolved	Conditional non-commercial only / legal review	Quarantine; public metadata only
Software Heritage archive (software_heritage)	Source-code archive	Possible legacy-language supplement; metadata-only	Not pinned (SNAPSHOT_TO_BE_SELECTED)	No	Unknown / record-level review — mixed per-origin/per-file terms	Unknown / record-level review	Public metadata only until snapshot, origin, and license resolution
The Stack v2 (stack_v2)	Repository metadata with Software Heritage content IDs	Metadata-only fallback	e565caa3a78c2423bd374333a472b049eb090e47	No; gated HF API metadata only	Unknown / record-level review — pinned card says license: other, requires adherence to original licenses and Software Heritage training principles	Unknown / record-level review — bulk access agreement, original licenses, attribution, and removal obligations apply	Public metadata fallback; no payload
RefineCode corpus metadata (refinecode_metadata)	Partial corpus metadata	Pipeline comparison; metadata-only	017558900665343ca563733212623e37606167ab	No; HF API metadata only	Unknown / legal review — pinned card reports MIT, while registry deliberately retains REVIEW_AT_PINNED_REVISION; underlying repositories are per-source	Unknown / legal review — full claimed payload is not public	Public metadata only
NVIDIA CUTLASS CuTe DSL (nvidia_cutlass_cute_dsl)	Python-embedded GPU DSL implementation/examples	Metadata-only; future domain adaptation	dcf215af68a2d08d305076c152a06f201728cd53	No; commit metadata only	Unknown / legal review — python/CuTeDSL is governed by an NVIDIA EULA, not the repository's top-level BSD-3-Clause license	Unknown / legal review — do not republish CuTe DSL files before clearance	Legal quarantine; public metadata/pointer only

Benchmark-family leakage firewall

These constraints apply even when a license would otherwise permit training:

• Build problem_family_id before translating, augmenting, or assigning a split. All prompts, paraphrases, language renderings, translation directions, reference solutions, tests, mutations, repairs, and completion surfaces of one problem are indivisible.
• Treat every HumanEval ID and all of its mHumanEval/MultiPL-E/other translations as one family. Treat MBPP derivatives likewise. MultiPL-E remains no-training regardless of family assignment.
• Treat all CodeNet submissions sharing a problem_id as one family. No language version of a problem may cross task training and evaluation.
• For SWE-rebench, group by repository plus PR/issue lineage and quarantine the leaderboard's complete task, repository, commit, PR, patch, and test families. Reconstructed base repositories are contamination evidence, not extra copies of raw pretraining data.
• McEval generation, explanation, and completion variants sharing a canonical problem remain together. McEval evaluation families must be removed from McEval-Instruct and raw repositories before any training payload is admitted.
• Run exact and near-match scans for benchmark prompts, tests, and solutions against all candidate raw, task, and SFT sources. On a cross-split match, evaluation quarantine wins and the entire connected repository/problem family is excluded from training.
• A benchmark used for training or prompt tuning cannot retain its standard-name evaluation score. Report it only as a separately labeled seen-family or training-overlap condition.

Release gate

payload/public metadata/quarantine in this document is a collection decision, not a completed rights determination. Public Code-ATLAS payload requires both training_allowed=true and redistribution_allowed=true at record level, preserved license/attribution data, global deduplication, benchmark-family decontamination, and an immutable source revision. Unknown or prohibited records remain visible only through source URL, revision, hashes when available, and exclusion reason.