dm27-v4m6

A generalist decision model for the REAX subnet (Bittensor SN92) generalist track, eval v5.1.

• Base: decisionmodels/reax-1.1 at 142b19355af00251f08db3d287307dd8b711c133; all credit for the base model and its lineage goes to its authors (see their model cards and NOTICE).
• Change: a LoRA fine-tune (merged into the weights) on the fixed REAX decision prompt, one question per prompt, letter readout, trained on procedurally generated decision tasks (arithmetic, causal reasoning, code execution, logic, rule application, games, tool selection) and permissively licensed public datasets (see the dataset list in the repository history of the training code).
• Serving: exactly as the base (same architecture and readout format).