README: contents, provenance and recorded LIBERO results

README.md

@@ -0,0 +1,67 @@
+---
+license: apache-2.0
+tags:
+  - robotics
+  - vla
+  - libero
+---
+
+# arc-layermap-ft
+
+Archived **EMA-only** weights for the layer-map LIBERO finetunes (`m323` protocol):
+the full-layer full-FT arm, the first-28 full-FT arm, and the first-28 frozen arm.
+Full checkpoint directories (32–44 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.step0000NNNNN/state.pt    EMA weights ONLY
+<run_dir>/<ckpt>.ckpt.shards.step0000NNNNN/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).
+
+Note the frozen run archives only its trainable set;
+the two full-FT runs carry the whole trunk.
+
+## Contents
+
+| run dir | step | EMA tensors | params | LIBERO | why this step |
+|---|---|---|---|---|---|
+| `lafm_vla_flow_neo2b_fulllayer_libero_fullft_m323_flft0815r` | `step000020000` | 1049 | 2.92 B | **94.80** | grid peak |
+| `lafm_vla_flow_neo2b_first28_libero_fullft_m323_f28ft0811` | `step000017500` | 737 | 2.13 B | **94.55** | grid peak |
+| `lafm_vla_flow_neo2b_first28_libero_fullft_m323_f28ft0811` | `step000020000` | 737 | 2.13 B | 94.25 | originally requested |
+| `lafm_vla_flow_neo2b_first28_libero_m323_f28ft0811` | `step000030000` | 376 | 0.44 B | **93.15** | grid peak and last step |
+
+Full 4-suite LIBERO (spatial / object / goal / long; 10 tasks x 50 trials x 4 =
+2000 rollouts per point), **m323** decode: replan 4, action horizon 8, NFE 10,
+no control_rate, `use_ema: true`, render 256, pinned mujoco 3.2.3 / robosuite 1.4.1.
+Single seed per point.
+
+**Do not compare these against the GEO or camera-mode archives** — those use the
+NA12 decode protocol (replan 12 / horizon 20 / control_rate 20) and the two
+protocols are not interchangeable.
+
+Two step choices deserve a note:
+
+- **first-28 full-FT** — the peak is 94.55 at **step000017500**, not at 20000
+  (94.25). Both are archived so the peak is not lost.
+- **first-28 frozen** — 93.15 at step000030000 is both the best and the last step.
+  An earlier internal note quoting "92.8" is stale: it is the rounded 92.75 at
+  step 5000, recorded while the grid still stopped at 17500.
+
+## Not included
+
+Only the steps above are archived; every other grid step was deleted and is **not**
+recoverable. Single-seed grid curves for the record (steps 2500 → 30000 by 2500):
+
+- fulllayer full-FT: 89.80 / 93.35 / 94.35 / 94.25 / 93.30 / 94.20 / 94.20 / **94.80** / 94.45 / 94.55 / *(27500 and 30000 were never evaluated)*
+- first-28 full-FT: 89.35 / 94.50 / 93.50 / 94.25 / 93.65 / 94.50 / **94.55** / 94.25 / 92.90 / 93.50 / 93.75 / 93.30
+- first-28 frozen: 85.80 / 92.75 / 91.75 / 92.45 / 92.45 / 92.15 / 91.85 / 92.65 / 92.05 / 92.25 / 92.25 / **93.15**