sn92-generalist (mammoth-1 joint-readout decider: our fine-tune F1 merged with two public entries)
A decision model for the REAX (Bittensor subnet 92) generalist track, 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). All parts share the base
baselabs/mammoth-1 (revision 03007ab3; revision 61cec44c has the same weight
files).
Recipe (details in MODIFICATIONS.md):
- D = the joint-readout conversion of mammoth-1 (readout rows initialised from its LM-head rows of the letter tokens).
- F1 = our fine-tune of D: one LoRA pass (rank 32) plus the readout, trained on decision questions in the REAX prompt. arXiv category and argument-stance questions take the answer distribution of our own earlier decider Andrew93110/sn92-gen-v1 as the target (on the questions it answers correctly); replay questions of every other kind keep D's own behaviour.
- This model = a task-arithmetic merge, tensor by tensor in float32:
F1 + 0.6 x (xflud - D) + 0.6 x (m87b - D), with xflud/sn92-decider-1 @ db2e622a and isomsom/sn92-gen-m87b @ 7f04058f (both MIT, both fine-tunes of mammoth-1 in the same layout). Their readout rows A..Z enter the merge the same way; the configuration and tokenizer files are F1's.
Credit: the merged changes of xflud/sn92-decider-1 and isomsom/sn92-gen-m87b are their authors' work; our own part is
F1 (step 2) and the choice of the merge. Data for F1: 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,
the merged models' retained notices and data attribution: NOTICE.md.
- Downloads last month
- 8