model card: 30-split SFT + eval

README.md

@@ -1,18 +1,12 @@
-# boxing — SLM-RL LoRA adapter
+# slm-rl-boxing — SFT LoRA (30-split bake)
 
 - **Game:** boxing
 - **Base model:** LiquidAI/LFM2.5-350M
-- **Champion generation:** 1
-- **Promoted:** True (baked pack adopted as RL initialization (not gated))
+- **Training:** reject_sft on 20 DQN teacher episodes (10 holdout reserved)
+- **Champion generation:** 1 (warm-start for GRPO / field E2E)
+- **Live gate eval (20 frozen seeds):** primary=-33.2 mean_score=-33.2 win_rate=0.0
+- **Smoke stream (seed 1000):** score -3 (base model same seed was -22)
 
-## Gate eval metrics
+Layout: `adapter/` PEFT files for SLM-RL `adapter_url` / playground paste.
 
-```json
-{
-  "note": "post-train eval skipped (--skip-eval)",
-  "primary": 0.0,
-  "skipped": true
-}
-```
-
-Trained with [SLM-RL](https://github.com/CraftsMan-Labs/SLM-RL) — a self-improving game gymnasium for small language models.
+Trained with [SLM-RL](https://github.com/CraftsMan-Labs/SLM-RL).