| # Data sources and release boundary |
|
|
| | Source or asset | Role | Terms recorded at the pinned revision | Release form | |
| |---|---|---|---| |
| | Kev executable generators and decision-v7 | TRAIN/replay | Apache-2.0 code; upstream dataset manifests vary | Builders, revisions and hashes; upstream acquisition where required | |
| | OpenJev executable-oracle fixtures | TRAIN/DEV/confirmation | OpenJev-authored deterministic fixtures | Source builders; original fixtures may be released under CC0-1.0 where marked | |
| | NVIDIA Aegis 2.0 | TRAIN/DEV | CC-BY-4.0; AART-derived rows excluded | Attribution plus source recipe | |
| | CodeReviewer | TRAIN/DEV | CC-BY-4.0 plus origin-repository terms | Recipe-only pending per-origin review | |
| | CommitPackFT | TRAIN/DEV | MIT dataset plus per-row repository license | Recipe-only pending row notices | |
| | FlakeFlagger | TRAIN/DEV | CC-BY-4.0/BSD-3-Clause plus project terms | Recipe-only pending project review | |
| | deepset/Lakera prompt injection | CALIBRATION | Apache-2.0 / MIT | Recipe and attribution; never TRAIN | |
| | When2Call | CALIBRATION | CC-BY-4.0 | Recipe and attribution; never TRAIN | |
| | BIG-Bench subset | DEVELOPMENT | Apache-2.0 repository, benchmark canary | Recipe-only; never TRAIN | |
| | LongBench-v2 | sealed LOCKED-TEST | Apache-2.0 metadata | Recipe-only, unopened for this selection | |
| | MMLU-Pro | sealed LOCKED-TEST | MIT data card / Apache-2.0 code | Recipe-only pending scope review | |
| | Belebele | sealed LOCKED-TEST | CC-BY-SA-4.0 | Recipe-only pending attribution/share-alike review | |
| | Qwen3.5-27B teacher responses | rejected R1 research arm | Apache-2.0 model; output ledger rights not assumed | Excluded; recipe and hashes only | |
| | Jev responses | evaluation-only | no redistribution assumption | Excluded; never used to train this model | |
|
|
| Core counts: |
|
|
| - K0 continuation: 8,000 local plus 8,000 deterministic Kev replay records. |
| - R1 TRAIN: 11,320 records / 16,279 questions. |
| - R1 DEVELOPMENT: 1,300 / 1,859. |
| - R1 CALIBRATION: 300 / 300. |
| - R4 TRAIN reused by later small-model research: 7,200 / 9,742. |
| - Capability-v1 consumed confirmation: 1,200 / 1,920. |
| - Confirmation-v2 consumed confirmation: 1,000 / 1,700. |
| - Sealed MMLU-Pro/Belebele design: 976 questions / 416 groups. |
|
|
| Dataset roles are immutable. DEVELOPMENT, CALIBRATION, consumed confirmation, |
| LOCKED-TEST, and Jev outputs cannot be converted into training data or mined to |
| construct new examples. Recipe-only assets are a rights boundary, not a claim |
| that their rows are OpenJev-owned. |
|
|
