File size: 2,505 Bytes
f1a00a5 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 | # 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.
|