v3 repair recipe
Base revision: 11e6f657f6d8830da230d4e49c65d0f67f106585. Seed: 2026101403.
- Apply documented legacy descriptor/contract decisions and quarantine unresolved source labels. Restore historical validation by canonical source ID with test precedence. Group normalized text, one leading greeting, and quoted/numeric surface forms against existing holdouts. Lower-priority rows are reserved; they are not relabeled or moved into a favorable evaluation split.
- Recover Hermes by exact first-user
[:400], target name and target-description[:160]matches. Validate original call contracts, preserve the full original request and native call set, and keep multi-call/singleton-tool records outside the choice view. Native request groups and historical siblings stay protected. - Regenerate every retained semantic present/absent pair with the repaired catalog and conservative conflict exclusions. Both variants have the same candidate count and exactly one ordinary candidate replacement. Native labels are not changed to conceal ambiguous requests.
- Exclude all steps of 31 reviewed bad tool parents. Recover Nemotron schemas from the original NVIDIA file; only strict verified prefixes plus valid native contracts enable argument supervision. Reserve exact policy/catalog environments from previously train-only groups; keep existing evaluation exposure metadata.
- Retain native NLU++ annotation and mask generic Boolean export. Remove only the invalid date values from the separate value-training view; retain original slots unchanged.
- Keep HINT3 native brand-relative labels and original test isolation. Apply the row-level finance/overlap/blank/duplicate decisions, split validation by query group inside native train only, and preserve original catalog semantics and licensing.
- Export deterministic random criterion IDs. Validate every choice record using the actual Clef encoder without truncation. Auxiliary tasks have no automatic choice conversion.
Auditable dispositions
The ledger has 2,524 entries: {"deduplicated":8,"quarantine":426,"separate_task":752,"split_reserved":1338}. These are disposition records, not an estimate of wrong labels. Examples retained in a native auxiliary view have separate_task; lexical reservations have split_reserved.
Scope of quality claims
Checks establish structural consistency, known-defect handling, recorded partition rules and source recovery. They do not establish exhaustive semantic correctness, original-dialogue isolation for all legacy sources, independent unseen behavior of pretrained models, or preserved model accuracy. Argument schema validity is not execution validity.
The current choice adapter uses published query text and candidate name/description only. Full schemas are retained separately for future reviewed adapters. HINT3, native NLU++ and Hermes call sets require their own objectives and evaluation. Do not count their rows or original label strings as new independent choice intents.