| # Permission audit and federated content |
|
|
| Permission is not a single Boolean. The registry records, for every canonical |
| source, separate findings for public access, local research use, model |
| training, commercial training, redistribution of original data, and |
| redistribution of transformed canonical trajectories. Each finding has both a |
| permission value and a verification value, plus source-specific conditions and |
| evidence URLs. |
|
|
| The audit vocabulary is: |
|
|
| - permission: `permitted`, `permitted_with_conditions`, `uncertain`, or |
| `restricted`; |
| - verification: `verified`, `inferred`, or `uncertain`; |
| - public payload: `hosted` or `metadata_only`. |
|
|
| The complete local inventory remains 155,278 trajectories. Publication does |
| not change scores, outcomes, ranks, splits, lineage or canonical hashes. |
| Instead, the public Hugging Face release contains the complete metadata index |
| and the payloads supported by the audit. Metadata-only sources retain pinned |
| upstream URLs and canonical shard hashes. |
|
|
| The initial public payload decision is: |
|
|
| - hosted: `swegym_openhands`, `nvidia_open_swe_traces`, `terminaltraj`, |
| `official_tau3`, `mind2web`, `toolbench`, |
| `swebench_verified_trajectories`, |
| `swebench_multilingual_trajectories`, `terminalbench_trajectories`, and |
| `swesmith_trajectories`; |
| - metadata-only: `agentsuite_tau2`, `swegym_sampled`, and |
| `browsecomp_rollouts`. |
|
|
| This is not a blanket declaration that hosted data is unrestricted or that |
| metadata-only data cannot be used. For example, Mind2Web combines CC-BY-4.0 |
| metadata with a research-purpose statement; tau benchmark training creates |
| evaluation contamination that must be disclosed; and software-engineering |
| records can embed code governed by per-repository licenses. |
|
|
| Downstream projects can select a policy directly: |
|
|
| ```python |
| from agent_trajectories import CanonicalDatasetAdapter, DatasetQuery |
| |
| dataset = CanonicalDatasetAdapter("./sft-benchmark-trajectories") |
| strict_research = DatasetQuery.for_training( |
| content_availability=("hosted",), |
| training_permissions=("permitted", "permitted_with_conditions"), |
| training_verifications=("verified",), |
| commercial_training_permissions=("permitted", "permitted_with_conditions"), |
| ) |
| for metadata in dataset.metadata(strict_research): |
| print(metadata.source_id, metadata.source_permissions.to_dict()) |
| ``` |
|
|
| An authorized user can explicitly ingest a metadata-only source at its pinned |
| revision and overlay the canonical shards: |
|
|
| ```python |
| dataset = CanonicalDatasetAdapter( |
| "./sft-benchmark-trajectories", |
| content_overlays={"agentsuite_tau2": "./canonical/agentsuite_tau2"}, |
| ) |
| ``` |
|
|
| The overlay is local and is never uploaded by the adapter. Canonical trajectory |
| IDs are matched against the persisted full-inventory index. Queries that select |
| missing content fail with `MissingTrajectoryContentError` and the relevant |
| source IDs instead of silently omitting records. |
|
|
