sn92-generalist (mammoth-1 joint-readout decider, F1)
A decision model for the REAX (Bittensor subnet 92) generalist track: a fine-tune of
baselabs/mammoth-1 (revision 03007ab3), stored in the REAX joint-readout
layout: a Qwen3_5Model backbone and readout.safetensors with answer rows A..Z, served through REAX's
full-bidirectional joint-readout protocol (non-causal full-attention layers, last-position pooling). mammoth-1 itself
was trained with block-bidirectional full-attention layers; REAX's mask is bidirectional over the whole prompt, an
approximation of that.
Recipe (details in MODIFICATIONS.md):
- the joint-readout conversion of mammoth-1 (readout rows initialised from its LM-head rows of the letter tokens);
- one LoRA pass (rank 32) plus the readout, trained on decision questions in the REAX prompt:
- arXiv category and argument-stance questions, with the answer distribution of our own earlier decider Andrew93110/sn92-gen-v1 as the target, on the questions it answers correctly (per-type distillation: that model is stronger on these two decision kinds);
- replay questions of every other kind, whose target is the converted model's own answer distribution, so its behaviour there stays in place; fact questions with an "I don't know" option learn to abstain where the converted model was wrong.
Data: arXiv API metadata (CC0), the args.me corpus (CC BY 4.0) and the replay sources listed in NOTICE.md. No
evaluation set of the REAX validator or of the Hugging Face / Artificial Analysis Decision Indices was used for training.
Licence: MIT (LICENSE). Qwen3.5-27B components remain under the Apache License 2.0 (LICENSE-APACHE-2.0). Lineage and
data attribution: NOTICE.md.
- Downloads last month
- 6