Expand description
The decision model: a frozen-architecture ModernBERT backbone plus the from-scratch
decision head (DecisionModel in rl_common.py).
Structs§
- Decision
Model - Outputs
- Raw model outputs for one batch: per-option logits (uncalibrated) and act-head probabilities.
Functions§
- stable_
softmax - Softmax in f32 with the max subtracted, matching the reference’s numpy path.