StatePlay mixed-SFT2000 RL no-filter checkpoints
This public repository contains two verified learned-parameter overlays requested on 2026-11-14. Both runs use Both mixed-dual reinforcement learning with group rejection disabled (keep_all). Each snapshot is frozen from its own verified latest.json pointer; the active training run was not stopped.
These files are parameter overlays rather than standalone models. Apply tensors by their exact state-dict names to the matching StatePlay mixed-SFT2000 base/runtime. The manifests record immutable source checkpoint and exported-weight SHA-256 values.
Snapshots
current-batch4-step00000148-cursor00001184— optimizer step 148, cursor 1184, 14,444,800 learned parameters.previous-no-filter-step00000320-cursor00002560— optimizer step 320, cursor 2560, 14,444,800 learned parameters.