README: final epoch-120 eval results (mean SR 86.1%, NJD 0.209)

README.md

@@ -26,7 +26,7 @@ diversity (NJD), computed via `kin_flow.cli.bench` on 10 held-out test scenes wi
 | `me-full_25000_5` | multi-embodiment | all 5 + z0 | 5 / 500 | 82.1% (mean) | 0.212 (mean) | MI300X (ROCm) |
 | `me-full_25000_30` | multi-embodiment | all 5 + z0 | 30 / 500 | 84.3% (mean) | 0.203 (mean) | MI300X (ROCm) |
 | `me-full_25000_90` | multi-embodiment | all 5 + z0 | 90 / 500 | 86.4% (mean) | 0.212 (mean) | MI300X (ROCm) |
-| `me-full_25000_120` | multi-embodiment | all 5 + z0 | **120** (target) | eval pending | eval pending | MI300X (ROCm) |
+| `me-full_25000_120` | multi-embodiment | all 5 + z0 | **120** (target) | 86.1% (mean) | 0.209 (mean) | MI300X (ROCm) |
 
 Single-embodiment models: `num_scenes=5000`. Multi-embodiment model: the paper's full
 configuration — `num_scenes=25000`, fp32, batch 5 scenes × 128 grasps, warmup-cosine LR
@@ -72,8 +72,20 @@ step; epoch numbers here are the corrected true dataset-pass count.
 Training finished 2026-07-15 at exactly **120.0 true dataset passes** (gradient step
 302,000 = 3,000,000 scene-visits / 25,000 scenes), the convergence point recommended by
 the original repo's README. Saved from the step-302,000 train-state checkpoint, so it
-also contains optimizer state for further fine-tuning. Per-gripper eval is pending and
-will be added here once complete.
+also contains optimizer state for further fine-tuning.
+
+| Gripper | DOF | SR | NJD |
+|---|---|---|---|
+| Panda | 2 | 97.5% | 0.257 |
+| VX300 | 2 | 97.4% | 0.134 |
+| DexEE | 12 | 73.0% | 0.114 |
+| Allegro | 16 | 85.4% | 0.288 |
+| Shadow Hand | 22 | 77.3% | 0.250 |
+| **mean** | | **86.1%** | **0.209** |
+
+Mean SR is flat vs. epoch 90 (86.1% vs 86.4%, within eval noise at 10 scenes × 100
+grasps), i.e. the model has converged — consistent with the original repo's guidance
+that ~120 epochs give reasonable convergence.
 
 ## Loading