win-reax-gen-1

A fine-tune of decisionmodels/reax-1 (revision ca2a0cc5) for the REAX SN92 generalist track (eval v5.1), in the base's joint-readout package format (protocol v3; readout.safetensors unchanged from the base).

Method. Three LoRA runs (rank 16 on all attention / linear-attention / MLP projections, lr 5e-5, up to 600 steps, readout head frozen), merged as the exact average of their weight deltas, then folded into the BF16 base weights. Training loss: cross-entropy on our own procedurally generated decision questions, plus a KL(base || model) penalty on public multiple-choice prompts using the base model's own answer distribution as the target (no labels from those sources were used in training).

Training data. Our own procedural generators (no REAX items, no dev-set items, no real personal data) and, for the KL anchors only, prompts from BoolQ (CC BY-SA 3.0), ARC (CC BY-SA 4.0), HellaSwag (MIT), GSM8K (MIT) and CRUXEval (MIT); see DATA_LICENCES.json.

Licence. Apache-2.0, as the base. NOTICE and LICENSE are the base's; our changes are listed in MODIFICATIONS.md.