taT (REAX SN92 generalist, weight-space merge)
A merge of 2 public fine-tunes of baselabs/mammoth-1 (03007ab3c4481bf01c71ef63adbbbc4b2093d71a, MIT; post-trained from Qwen3.5-27B, Apache-2.0) in the REAX joint-readout layout (Qwen3_5Model backbone, readout.safetensors, decision_config.json). No training of our own: all learned changes come from the parents below, whose authors deserve the credit.
Parents
| Model | Revision | Licence | Weight of its change |
|---|---|---|---|
| isomsom/sn92-gen-m87b | 7f04058f63745440fae3c10f8da249a0c55fea40 |
MIT | 1 |
| xflud/sn92-decider-1 | db2e622aa508bd82350620bf165441fa4043f3cf |
MIT | 1 |
Recipe
For every backbone tensor, with base = baselabs/mammoth-1@03007ab3c4481bf01c71ef63adbbbc4b2093d71a: TIES (per tensor: keep each parent's change top-0.5 entries by magnitude, elect the sign by the weighted sum, average the agreeing entries) with weights m87b 1, dec1 1, lam 1; computed one tensor at a time in float32 and rounded once to bfloat16. Readout: the weighted average of the parents' readouts after dividing each by its least-squares scale against the base readout. Config, tokenizer and decision/joint configs are the base's. MERGE.json records the recipe.
Licence
MIT (LICENSE keeps the copyright lines of every parent); NOTICE retains each parent's notice verbatim; Qwen3.5-27B terms in LICENSE-APACHE-2.0.
- Downloads last month
- 9