Document recovered Spa-Bench training provenance
Browse files
README.md
CHANGED
|
@@ -52,6 +52,12 @@ revision. The current public GR00T base revision at archival review is
|
|
| 52 |
`2fc962b973bccdd5d8ce4f67cc63b264d6886495`; it must not be assumed to be the
|
| 53 |
unrecorded training revision.
|
| 54 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 55 |
## Deployment processing
|
| 56 |
|
| 57 |
For both GR00T variants, non-gripper action dimensions used the causal filter
|
|
|
|
| 52 |
`2fc962b973bccdd5d8ce4f67cc63b264d6886495`; it must not be assumed to be the
|
| 53 |
unrecorded training revision.
|
| 54 |
|
| 55 |
+
The author-supplied original
|
| 56 |
+
[`train_groot_n17_frozen_llm.sh`](https://github.com/justintiensmith/Imperial-Thesis/blob/main/training/groot/train_groot_n17_frozen_llm.sh)
|
| 57 |
+
launcher is archived with the thesis artifact. It records seed 42, four-GPU
|
| 58 |
+
training, the immutable dataset revision, component-freezing flags, checkpoint
|
| 59 |
+
cadence, and the historical environment paths used for this run.
|
| 60 |
+
|
| 61 |
## Deployment processing
|
| 62 |
|
| 63 |
For both GR00T variants, non-gripper action dimensions used the causal filter
|