StatePlay Mixed SF3 + KOF98 Visual-Only RL
Public weight snapshot from the single-stage visual-only RL experiment. It was initialized independently from the shared Base SFT and updates visual LoRA only.
- Optimizer step: 58
- Completed condition cursor: 656
- Trainable parameter union: 47,185,920 parameters / 480 tensors
- Reward:
0.6 * mechanics + 0.4 * visual_quality - Mechanism guard filtering: enabled
- Source SFT SHA256:
d22492703ed233088108663955c2360b874f6c32dfe99ae18fe19dd8d4ab7077 - Frozen training checkpoint SHA256:
d85c26977f651eef7a8b9bb8c99079c1df56914531a95a94768aecb2ff034df3
model.safetensors is a weights-only learned-parameter union. It does not
contain optimizer or RNG state. Load it together with the matching base SFT and
transition_tokenizer.pt; tensor names are preserved exactly.