isomsom/sn92-gen-v6

REAX SN92 generalist entry

Decision model for the REAX (SN92) generalist track, served with the validator's fixed prompt and letter readout (softmax at T=2).

Lineage and modifications (Apache-2.0 sec. 4)

Qwen/Qwen3.8-27B -> vllm-sr/d3@cf15549a78fc52598ddd1951f321c58372bbd64f (Apache-2.0, pplx-format decision model) -> rewritten to the joint readout layout (decision_config v2) -> LoRA r32 trained in joint mode (bidirectional full attention, as the REAX validator serves it). Stage 1 (lr 5e-5, 0.5 epoch): public SN92 specialist train rows, public benchmark TRAIN splits and our own written items. Stage 2 (lr 5e-5, 531 steps of 32 rows), continued from stage 1: public SN92 specialist train, factory and dev rows (decisionmodels/sn92-specialists, CC-BY-4.0) plus the same extras. Stage 3 (lr 5e-5, 600 steps of 32 rows, another data order), continued from stage 2: the same mix plus the remaining public SN92 specialist train rows and the second factory half (decisionmodels/sn92-specialists, CC-BY-4.0); merged into the backbone. Stage 4 (a new LoRA r32 on that merged model, lr 3e-5, 300 steps of 32 rows): replay of the same mix plus public procedural TRAIN rows (tasksource/procedural-typed-decisions, Apache-2.0). Merged into the backbone; readout rows A..Z retrained and rescaled for the fixed T=2 readout.

Declared base for the track: Qwen/Qwen3.8-27B@1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0. Weights only; no code. Licence: Apache-2.0 (see LICENSE).