Document private checkpoint and data artifacts
Browse files
README.md
ADDED
|
@@ -0,0 +1,48 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: cc-by-nc-sa-4.0
|
| 3 |
+
task_categories:
|
| 4 |
+
- robotics
|
| 5 |
+
tags:
|
| 6 |
+
- robotwin
|
| 7 |
+
- lerobot
|
| 8 |
+
- posttraining
|
| 9 |
+
---
|
| 10 |
+
|
| 11 |
+
# SWAM posttraining data
|
| 12 |
+
|
| 13 |
+
RoboTwin Clean data from the embodied-model storage used by the selected SWAM 12k posttraining run, packaged as one TAR archive per task.
|
| 14 |
+
|
| 15 |
+
- 50 tasks, 2,500 stored episodes, and 551,931 stored frames.
|
| 16 |
+
- 16 excluded episodes are recorded in the original per-task exclusion manifests; applying those manifests leaves 2,484 eligible episodes.
|
| 17 |
+
- Original episode IDs, Parquet files, videos, instructions, statistics, and metadata are preserved.
|
| 18 |
+
- **Latents are not included.** No augmented dataset is included.
|
| 19 |
+
- The stored source files total 3,922,419,040 bytes across 10,814 files. TAR sizes include archive overhead.
|
| 20 |
+
|
| 21 |
+
## Layout and extraction
|
| 22 |
+
|
| 23 |
+
`data/<task-directory>.tar` contains `<task-directory>/data/`, `<task-directory>/videos/`, and `<task-directory>/meta/`.
|
| 24 |
+
|
| 25 |
+
Extract all archives into a common directory named `lerobot_robotwin_eef_clean_50` to restore the task directory structure. For example:
|
| 26 |
+
|
| 27 |
+
```bash
|
| 28 |
+
mkdir -p lerobot_robotwin_eef_clean_50
|
| 29 |
+
for archive in data/*.tar; do
|
| 30 |
+
tar -xf "$archive" -C lerobot_robotwin_eef_clean_50
|
| 31 |
+
done
|
| 32 |
+
```
|
| 33 |
+
|
| 34 |
+
The training instructions are `meta/tasks_v1.jsonl`. The training exclusion manifest is `meta/episode_exclusions_20261023.jsonl`. **Excluded episodes remain in the archives** to preserve the original dataset and IDs. Consumers must apply the exclusion manifests; some excluded episodes have known video decoding failures or manually identified action anomalies.
|
| 35 |
+
|
| 36 |
+
## Integrity
|
| 37 |
+
|
| 38 |
+
- `tasks.json`: per-task episode/frame counts, exclusion counts, and source file totals.
|
| 39 |
+
- `file-manifest.json`: SHA-256 and size of every archived source file.
|
| 40 |
+
- `archive-manifest.json`: SHA-256 and size of every TAR archive.
|
| 41 |
+
|
| 42 |
+
Every archive member was compared against its source file by SHA-256 during packaging. Uploaded archive hashes are verified against Hugging Face metadata.
|
| 43 |
+
|
| 44 |
+
## Provenance and license
|
| 45 |
+
|
| 46 |
+
The original dataset README is preserved as `SOURCE_README.md`; it identifies the source data license as [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/). Its reference to extracted latents describes the upstream source; those latents are intentionally omitted here. The upstream README refers to a LICENSE file that was not present in the selected source directory.
|
| 47 |
+
|
| 48 |
+
Paired checkpoints: [FreDERicia/SWAM](https://huggingface.co/FreDERicia/SWAM).
|