README: contents, provenance and recorded LIBERO results

README.md

@@ -0,0 +1,56 @@
+---
+license: apache-2.0
+tags:
+  - robotics
+  - vla
+  - libero
+---
+
+# arc-camode-ft
+
+Archived **EMA-only** weights for the six camera-mode (`camode`) LIBERO finetunes —
+the geoA and geoB backbones each finetuned under three camera regimes: `random`,
+`side`, `wrist`. Full checkpoint directories (~46 GiB/step, optimizer state
+included) were deleted from the cluster after this upload verified.
+
+## What each file is
+
+```
+<run_dir>/config.yaml                                  training config
+<run_dir>/dataset_stats.json                           normalizer stats
+<run_dir>/<ckpt>.ckpt.shards.step000020000/state.pt    EMA weights ONLY
+<run_dir>/<ckpt>.ckpt.shards.step000020000/engine_meta.pt
+```
+
+`state.pt` holds `{"extra": {"ema": {name -> fp32 CPU tensor}}}`, the payload
+`methods/lafm/vla_flow_ft.py:load_pretrain_trainables()` reads. Load with
+**`model.init_use_ema=true`**. No `"model"` key, no optimizer state — **init / eval
+only, not resumable**. Each `engine_meta.pt` is the original and its
+`ema_fingerprint` matches the archived tensors (rel < 1e-6).
+
+## Contents — all at `step000020000`
+
+| run dir | arm | camera mode | off-mode | with-R |
+|---|---|---|---|---|
+| `lafm_vla_flow_neo2b_camArft_camAr0816` | A | random | 12.10 | **97.65** |
+| `lafm_vla_flow_neo2b_camAsft_camAs0817` | A | side   | not run | **97.50** |
+| `lafm_vla_flow_neo2b_camAwft_camAw0817` | A | wrist  | not run | **98.20** |
+| `lafm_vla_flow_neo2b_camBrft_camBr0816` | B | random | 95.75 | **97.75** |
+| `lafm_vla_flow_neo2b_camBsft_camBs0817` | B | side   | 95.75 | **98.50** |
+| `lafm_vla_flow_neo2b_camBwft_camBw0817` | B | wrist  | 95.60 | **98.35** |
+
+Full 4-suite LIBERO (spatial / object / goal / long; 10 tasks x 50 trials x 4 =
+2000 rollouts per cell), **NA12** decode: replan 12, horizon 20, NFE 10,
+control_rate 20, `use_ema: true`, render 256, pinned mujoco 3.2.3 / robosuite 1.4.1.
+Single seed per cell.
+
+**Reading the two columns.** "with-R" supplies the camera-mode conditioning at eval;
+"off-mode" withholds it. Arm A collapses without conditioning (12.10) because arm A
+has no text fallback path — that is by design, not a regression. Arm B degrades
+gracefully (~95.6–95.8). Off-mode was never run for camA side/wrist.
+
+## Not included
+
+Only `step000020000` is archived per run. Steps 2500–17500 were deleted and are
+**not** recoverable. Two extra evaluated points that are now gone:
+camArft @15000 off-mode 14.35, camBrft @15000 off-mode 95.55.