Public repository with automatic gated access, preserving the access setting of the source SFT dataset. These are research artifacts from an unsuccessful identity-transfer campaign, not evidence of an effective training recipe.

used-prompts contains the exact 8,590-row pool packaged in data version 2026.10.23.5 for the ratio and masking ablations. candidate-prompts contains a later 8,666-row expansion; its preparation is not evidence of use in those ablations. They overlap and must not be concatenated. Both contain student generation contexts, not supervised target answers; thinking modes are balanced exactly. Route counts do not define sampling weights. The ratio sweep used 0, 16, 32, 48 or 64 identity trajectories per 64, with remaining trajectories using the base teacher. The masking arms used 32/32. Follow-up contexts supply multi-turn coverage.

corrected-validation has 212 references: 62 identity references from the released SFT teacher's excluded training split, 99 archived base-model general responses, and 51 successful policy-evaluation initial SWE actions. Identity reasoning is curated SFT text; base-model reasoning is generated. This is development validation, not an untouched final benchmark. The SWE slice does not measure task success. Private source-provenance metadata was omitted; token IDs, reference boundaries, text and other scoring fields are unchanged. Do not train on this split or its paraphrases when using it for evaluation.

The identity prompt lineage is the V13/V19 SFT dataset; additions were reviewed by Codex, not independent humans. General preservation contexts and SWE initial actions derive from the campaign's v0.1 evaluations. This publication does not grant a new blanket license over source benchmark material; consult the original sources for reuse terms. Manifests record exact training-file hashes and validation transformation. JSONL copies are provided for inspection.

See results and report and experiment issue. Aggregate reference NLL improved without direct identity accuracy improving. Validation scores and reviewer limitations are documented in the report.

Downloads last month: 24

Total file size:

11 MB