REAX SN92 generalist entry (merge)
A weight-space merge of three public REAX SN92 generalist entries that share the base
baselabs/mammoth-1 (61cec44c281c693992e3d6ebeb44dd3402dfa186). No training of our own: all
learned changes come from the parents below, whose authors deserve the credit for them.
Parents
| Model | Revision | Weight of its change |
|---|---|---|
| isomsom/sn92-gen-m87b | 7f04058f63745440fae3c10f8da249a0c55fea40 |
0.5 |
| xflud/sn92-decider-1 | db2e622aa508bd82350620bf165441fa4043f3cf |
1 |
| edinlong83/sn92-generalist-h5 | 764dad666e3018c4724c02cc43d6e882f8dd431b |
0.5 |
Recipe
Task arithmetic: for every tensor the base also has, base + 0.5 * (parent_1 - base) + 1 * (parent_2 - base) + 0.5 * (parent_3 - base) with base =
baselabs/mammoth-1@61cec44c281c693992e3d6ebeb44dd3402dfa186 (its names carry a leading model.), i.e. parent i keeps weight_i of its own change from the base
(isomsom/sn92-gen-m87b: 0.5, xflud/sn92-decider-1: 1, edinlong83/sn92-generalist-h5: 0.5). The answer readout (readout.safetensors, not in the
base) is the average of the parents' readouts with the same weights, renormalised to sum 1.
Computed one tensor at a time in float32 and rounded back to bfloat16 (round to nearest even). All parents have
identical tensor names, shapes and dtypes (Qwen3_5Model joint-readout package).
config.json, decision_config.json, joint_config.json and the tokenizer are copied unchanged from the first parent.
Licence
MIT. See LICENSE (copyright lines of every parent retained) and NOTICE (each parent's notice retained verbatim). This model is not affiliated with, endorsed by, or sponsored by Base Labs / Baseten, the Qwen team, Perplexity AI or the parents' authors.
- Downloads last month
- 14