README: eval results (T009 pass 1), 15000/ and cont15k lineage

README.md

@@ -10,7 +10,19 @@ tags:
 # pi05_atomic65_from_base
 
 A π0.5 checkpoint (openpi, JAX) fine-tuned from `pi05_base` on **RoboCasa365 atomic tasks only**. It contains no composite-task data.
-This is a research checkpoint and has not been evaluated yet.
+This is a research checkpoint.
+
+## Evaluation (RoboCasa365, 65 atomic tasks, pretrain split, 5 episodes per task)
+
+openpi `eval_g0` client, replan every 5 steps, seeds `700000 + episode`, episodes 0–4. Mean success rate over the 65 tasks:
+
+| Checkpoint | Mean SR |
+|---|---|
+| `9999/` (10k) | 52.0% |
+| `15000/` (15k) | 62.5% |
+| RoboCasa365 `pi05_pretrain_human300` 75k (reference, trained on 300 tasks) | 40.3% |
+
+With 5 episodes per task the per-task numbers are noisy; the paired differences (same seeds) are significant (McNemar p < 0.001 for 15k vs 10k and for 10k vs 75k).
 
 | | |
 |---|---|
@@ -20,24 +32,28 @@ This is a research checkpoint and has not been evaluated yet.
 | Data | 65 RoboCasa365 atomic tasks: `get_ds_soup(split="pretrain", task_set="all_atomic_tasks", source="human")`, i.e. human demos, pretrain split, filter key `100_demos` (~100 demos per task) |
 | Norm stats | combined from each dataset's `meta/stats.json` (`assets/norm_stats.json`) |
 | Batch | 64 (2×A100 80GB, FSDP 2) |
-| LR | cosine: warmup 1k, peak 2.5e-5, decay over 100k steps to 2.5e-6. The run stops at 10k steps, so the LR is still near peak (≈2.49e-5 at 5k, ≈2.45e-5 at 10k) |
+| LR (steps 0–10k, config `pi05_atomic65_from_base`) | cosine: warmup 1k, peak 2.5e-5, decay over 100k steps to 2.5e-6. The run stops at 10k steps, so the LR is still near peak (≈2.49e-5 at 5k, ≈2.45e-5 at 10k) |
+| LR (10k–15k, config `pi05_atomic65_cont10k`) | init from `9999/` params, fresh optimizer; warmup 500, cosine planned over 20k steps from 2.5e-5 to 2.5e-6, stopped after 5k steps (LR ≈2.2e-5) |
+| LR (15k–50k, config `pi05_atomic65_cont15k`) | init from `15000/` params, fresh optimizer; warmup 500, cosine over 35k steps from 2.5e-5 to 2.5e-6 (in progress) |
 
 ## Layout
 
 | Folder | Train step | Contents |
 |---|---|---|
 | `5000/` | 5,000 | `params/` (orbax OCDBT), `assets/norm_stats.json` |
-| `9999/` | 10,000 (openpi saves the last step as `num_train_steps - 1`); added when training finishes | `params/` (orbax OCDBT), `assets/norm_stats.json` |
+| `9999/` | 10,000 (openpi saves the last step as `num_train_steps - 1`) | `params/` (orbax OCDBT), `assets/norm_stats.json` |
+| `15000/` | 15,000 (`pi05_atomic65_cont10k` step 5000) | same |
+| `20000/` … `45000/`, `49999/` | 15,000 + `pi05_atomic65_cont15k` step; added every 5k steps while training runs | same |
 
 The optimizer state (`train_state`) is not included.
 
 ## Usage
 
-Download one step folder and serve it with openpi:
+Download one step folder and serve it with openpi (all folders share the model and norm stats, so any of the three configs works for serving):
 
 ```bash
-hf download ducido/pi05_atomic65_from_base --include "5000/*" --local-dir ckpt
-uv run scripts/serve_policy.py policy:checkpoint --policy.config=pi05_atomic65_from_base --policy.dir=ckpt/5000
+hf download ducido/pi05_atomic65_from_base --include "15000/*" --local-dir ckpt
+uv run scripts/serve_policy.py policy:checkpoint --policy.config=pi05_atomic65_from_base --policy.dir=ckpt/15000
 ```
 
 ## License