sn92-generalist-s6

A LoRA fine-tune of baselabs/mammoth-1, packaged as a REAX joint-readout decider for the REAX (Bittensor subnet 92) generalist decision track. This repository contains weights and configuration files only: no code and no training data.

Base and format

item	value
declared base	baselabs/mammoth-1@61cec44c281c693992e3d6ebeb44dd3402dfa186 (MIT)
base lineage (its card)	Qwen/Qwen3.5-27B (Apache-2.0)
architecture	Qwen3_5Model backbone + readout.safetensors (one row per answer code)
decision format	decision_config.json format 2, noncausal_full_attention, last-position pooling; joint_config.json joint-question inline readout v1

mammoth-1 is a decision model trained with bidirectional full-attention layers. This package exposes it through the REAX joint-readout runtime (JointRunner), which serves it with the full-attention mask it was trained for. The readout rows start from mammoth-1's language-model-head rows for the answer-code tokens.

Method

1. Conversion: mammoth-1's language-model weights mapped one-to-one onto the Qwen3_5Model layout (same tensors, same values); the language-model head is not shipped, its answer-code rows become readout.safetensors.
2. Fine-tune: one LoRA pass (rank 64, alpha 128, attention and MLP projections of all layers, learning rate 3e-5 with warm-up and decay, about 4.9M tokens, readout frozen), cross-entropy on the gold answer plus a KL anchor (weight 0.1) to the converted start model.
3. Release: the LoRA delta merged into the backbone at weight 0.75, stored in bf16.

Run it like any REAX joint-readout decider; config.json has no auto_map.

Training data credits

About 40 % computed items from our own procedural generators (multi-step logic, arithmetic word problems, code output, rule decisions, tool-call and API decisions, causal yes/no, chords from notes), with gold answers computed by code.

About 35 % decision items built by us from public records:

• GitHub Security Advisories (GHSA, CC-BY-4.0): ecosystem, severity and CWE decisions;
• U.S. FDA enforcement reports (openFDA, public domain): classification, status and root-cause decisions;
• U.S. Federal Register documents (public domain): document type and agency decisions;
• OpenAlex works (CC0): research-domain decisions.

About 25 % replay of public decision data:

• decisionmodels/sn92-generalist-train-v1 and decisionmodels/sn92-specialists, credited per their cards;
• items derived from glaiveai/glaive-function-calling, NousResearch/hermes-function-calling, hotpotqa/hotpot_qa, BeIR/fever, coastalcph/lex_glue, deepmind/aqua_rat, allenai/ai2_arc, tau/commonsense_qa, openlifescienceai/medmcqa, google/boolq, Davlan/sib200, klue/klue, e9t/nsmc, tblard/allocine, zeroshot/twitter-financial-news-sentiment and Amazon ESCI, each under its own licence.

Licence

MIT, as the declared base. See LICENSE and NOTICE.

This model is not affiliated with, endorsed by, or sponsored by baselabs, the Qwen team or the REAX subnet owners.