Do not use this revision as a semantic model or as a training parent. A post-publication audit proved that every initial F7 stage rebuilt a new BPE tokenizer while inheriting weights whose embedding and output rows were indexed by the preceding tokenizer. The checkpoint is retained only as reproducible negative evidence.
7b24eb73442400856d4306c611e834e0fdab61e713f57ff9a5d6f88e8f80b699.
Its embedded tokenizer has order-independent token-mapping SHA-256
eb8a5347f5a88cdca2ce26ca1f5748dfd2180f952ccdba374e7d1c948bca6651,
whereas the last valid parent checkpoint used token-mapping SHA-256
25c0085f847e5304ddfc3e031f747ecd1d2b291f284ff559724bdb80d0f7ca08.
This changed token IDs beneath inherited parameters and invalidates the
semantic lineage even though the run was numerically finite and its held-out
loss decreased.b4b75f28b47b9b75d4d50e6a631c0a745cea6c55. The complete original training
artifact repository is also retained as negative evidence, not as an admitted
model.