# InternW0 LIBERO checkpoint provenance

## Selected artifact

`InternW0-Libero.pt` is the model-only checkpoint selected from LIBERO run
`w0-lib-gl-rf-c16-e10-32g-r13` at step 600.

Step 600 was selected over the final step 660 because it had the highest
complete four-suite evaluation average and the largest number of successful
rollouts:

| Checkpoint | Spatial | Object | Goal | Long | Average | Successes |
|---|---:|---:|---:|---:|---:|---:|
| Step 600 (selected) | 99.4% | 99.4% | 98.6% | 97.0% | 98.6% | 1,972 / 2,000 |
| Step 660 | 99.0% | 99.4% | 98.4% | 97.2% | 98.5% | 1,970 / 2,000 |

## Lineage

1. Heterogeneous InternW0 pretraining at global step 30,000.
2. LIBERO target-domain initialization with `new_domain` action/proprio
   interfaces, seed 42.
3. Earlier LIBERO C660 run, selected at step 2,640.
4. R13 Goal+Long run, initialized from the step-2,640 LIBERO checkpoint and
   selected at step 600.

The historical source implementation was `ahawam`; the public runtime is `w0`.
The source reader and public `w0` reader have identical Python ASTs apart from
non-semantic comments. The resolved training config is preserved in
`source_training_config.yaml`.

The checkpoint embeds `behavior_config`, `training_recipe`,
`posttrain_statistics`, and `posttrain_provenance`. These fields are exported
verbatim to `checkpoint_metadata.json`.

## Training contract

- LIBERO suites used for posttraining: `libero_goal` and `libero_10`.
- Random-frame sampling with 100,000 samples per epoch.
- 10 epochs.
- Batch size 48 per process.
- Learning rate 5e-5, cosine schedule.
- Seed 42.
- Native 20 Hz EEF10 representation.
- 37-dimensional canonical action/proprio interface.
- Native EEF10 slots:
  `[7, 8, 9, 10, 11, 12, 13, 14, 15, 16]`.
- Action horizon 64, chunk size 16, video plan horizon 96, video/action stride 8.
- Maximum action offset 15.
- K=8 target soft prompts.

The matching normalization file is `normalization.json`.

## Evaluation record

The complete evaluation uses all 40 tasks in the four official LIBERO suites,
50 trials per task, for 2,000 stochastic rollouts. The evaluation uses:

- environment/init seed 42;
- unset policy diffusion seed;
- 30 settling steps with `[0, 0, 0, 0, 0, 0, -1]`;
- execution of all 16 actions per predicted chunk;
- 10 inference steps;
- runtime gripper binarization;
- lossless PNG image transport;
- 400-step limits for Spatial/Object/Goal and 700 for `libero_10`;
- MuJoCo 3.3.2.

The full per-task record is included in `evaluation_summary.json`, and its
launch metadata is included in `evaluation_meta.json`.

## Excluded from this repository

The checkpoint package does not include optimizer state, distributed training
state, logs, LIBERO data, Wan2.2 files, DINOv3 weights, text caches, or the
external LIBERO checkout.