Upload tinyvla2/CLAUDE.md with huggingface_hub

tinyvla2/CLAUDE.md

@@ -191,9 +191,11 @@ Currently training: `tv2_libero_ft` — C-scaled fine-tuned on LIBERO in canonic
 action space with morphology conditioning, 30k steps, to get a closed-loop number
 directly comparable to v1's 86%. At 5k it scores 40% on libero_spatial.
 
-`tv2_robocasa` was stopped at ~18k/30k to free the GPU; its 10k checkpoint gave
-RoboCasa 0.52 (best-in-project on that domain) but degraded held-out LeKiwi, i.e.
-weight 0.5 is too high for the new domain.
+`tv2_robocasa` was stopped at ~18k/30k to free the GPU; only `step_10000` was kept.
+Re-measured with the corrected harness: RoboCasa (new domain) **0.47**, at the cost
+of mild uniform forgetting on the old robots (mean 0.34 → 0.41, worst stretch
+0.11 → 0.25) and **no** effect on held-out LeKiwi (1.67 → 1.60). The earlier
+"weight 0.5 collapses held-out transfer" conclusion was a harness artefact.
 
 Benchmark commands:
 ```bash