legacy: add index of archived experimental checkpoints

legacy/README.md

@@ -0,0 +1,109 @@
+# `legacy/` — archived experimental checkpoints
+
+Superseded runs, kept for reference. Archived **2026-08-11**; nothing here is deleted,
+only moved out of the repo root so the active models are easy to find.
+
+**These are exploratory spatial-branch experiments, mostly single-task (task 12 / task 31
+boxing gloves / clean-up-your-desk).** For anything you actually want to run, use the
+folders at the repo root instead — see [Active models](#active-models-at-the-repo-root).
+
+Total: **13 folders, 793 files, 173.6 GiB.**
+
+---
+
+## What is in here
+
+### `kvsplit_*` — K/V-split DA3 bank
+
+The DA3 spatial-bank redesign where bank **keys** carry only the address (position +
+Plücker ray + view index) and **values** carry only the payload, instead of both being
+projections of the same concatenated feature.
+
+| folder | size |
+|---|---|
+| `kvsplit_newbank_step39999` | 14.10 GiB |
+| `kvsplit_spatretrain_step12000` | 13.23 GiB |
+| `kvsplit_spatretrain_step18000` | 13.23 GiB |
+| `kvsplit_spatretrain_step19999` | 13.23 GiB |
+| `kvsplit_strongbase_step14999` | 13.23 GiB |
+
+`spatretrain` is one run captured at three steps — 12000, 18000 and 19999. Per its
+original commit message, **step 18000 was the best of the three** (loss ≈ 0.115), so 19999
+is not automatically the one to reach for.
+
+### `task12_da3_*` — task-12 DA3 ablations
+
+Same task, varying the depth backbone (`giant` vs `large`), the geometry representation
+(`gtdepth` vs `pointmap`), and the LR treatment of the base model (`1p5x`, `slowbase`).
+
+| folder | size |
+|---|---|
+| `task12_da3_giant_gtdepth_v2_1p5x_step19999` | 13.33 GiB |
+| `task12_da3_large_gtdepth_newbank_step19999` | 13.32 GiB |
+| `task12_da3_large_gtdepth_v2_1p5x_slowbase_step19999` | 13.33 GiB |
+| `task12_da3_large_gtdepth_v2_pointmap_step19999` | 13.32 GiB |
+| `task12_da3_large_gtdepth_v2_pointmap_step29999_ext30k` | 13.32 GiB |
+
+`..._pointmap_step29999_ext30k` is the 30k-step extension of the 20k `..._pointmap_step19999`
+run, not an independent experiment.
+
+### `vggt_*` — VGGT spatial backbone
+
+The alternative to DA3: VGGT as the geometry encoder, on the boxing-gloves task.
+
+| folder | size |
+|---|---|
+| `vggt_newbank_boxing_gloves_step19999` | 13.34 GiB |
+| `vggt_newbank_boxing_gloves_step8000` | 13.34 GiB |
+
+### `pi05_da3_freezebase_desk_step8000` — 13.23 GiB
+
+DA3 with the base policy frozen, on `clean_up_your_desk`. An early spatial-branch probe.
+
+---
+
+## Active models, at the repo root
+
+| folder | what |
+|---|---|
+| `meta5_2026_224/` | 5-task meta-training — the root `README.md` documents this one |
+| `meta5_2026_224_ext/` | extension of the above past 20k steps |
+| `task77_installing_a_modem/` | single-task fine-tune, with `README.md` + `EVAL.md` |
+| `task85_putting_dirty_dishes_in_sink/` | single-task fine-tune, with `README.md` + `EVAL.md` |
+| `single-task-finetune-0807/` | five single-task fine-tunes: `turning_on_radio`, `setting_mousetraps`, `make_microwave_popcorn`, `dispose_of_glass`, `installing_a_modem` |
+| `task31_clean_boxing_gloves_pi05_baseline_20k_step19999_params_only/` | no-DA3 baseline for the boxing-gloves task — the control the `vggt_*` runs in here were compared against |
+| `pibehavior_da3_clean_up_your_desk_40k/` | 40k-step DA3 run on `clean_up_your_desk` |
+| `bringing_in_wood_20k/`, `can_meat_20k/`, `can_meat_30k/`, `chop_an_onion_20k/`, `clean_up_your_desk_20k/` | earlier single-task runs |
+| `da3-newbank/` | DA3 spatial-branch weights only (0.8 GiB), not a full policy |
+| `norm-stats-fixed/` | standalone corrected `norm_stats.json` (robot-frame `base_qvel`) |
+
+`EVAL.md` (identical in `task77_*` and `task85_*`) is the shared eval contract — observation
+layout, wrapper behaviour, the full loop — and applies to every checkpoint in the repo.
+
+---
+
+## Before you use anything in here
+
+1. **Use each checkpoint's own bundled `assets/`.** Every folder ships its own
+   `assets/IliaLarchenko/behavior_224_rgb/norm_stats.json` and `fast_tokenizer/`. Do not
+   share one set across models, and do not substitute `norm-stats-fixed/` into these — all
+   of these runs predate the `base_qvel` correction and their weights adapted to the 2025
+   world-frame statistics. Swapping in the corrected file would put state dims 0:3 roughly
+   an order of magnitude off the scale they were trained on.
+
+2. **Loss values are not comparable across folders.** Different tasks, different step
+   counts, and episode length varies several-fold between tasks.
+
+3. **These need the DA3 (or VGGT) spatial branch wired in** — they are not stock
+   PiBehavior and will not load into a plain π0.5 eval path.
+
+4. **Training-set metrics only.** No validation split, no rollout evaluation.
+
+---
+
+## Note on how these were moved
+
+Moved with the Hub API, not re-uploaded: LFS blobs were copied server-side by reference,
+so the 173.6 GiB never left the Hub. Every one of the 793 files was verified identical at
+the destination (LFS `sha256` for params, git blob id for the rest) *before* the originals
+were deleted. Original upload dates are recoverable from the repo's commit history.