Add pretraining initialization weights
Browse files- pretrain/README.md +7 -0
- pretrain/SHA256SUMS +1 -0
- pretrain/manifest.json +9 -0
- pretrain/model.pt +3 -0
pretrain/README.md
ADDED
|
@@ -0,0 +1,7 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Pretraining weights
|
| 2 |
+
|
| 3 |
+
Weights-only initialization for main-policy fine-tuning. The action output head is omitted; robot-specific input and output projections are initialized by the training loader. This checkpoint is not intended for direct policy inference.
|
| 4 |
+
|
| 5 |
+
Use the standalone training loader with `--init-weights path/to/pretrain/model.pt`. The root repository's `config.yaml` and `dataset_stats.json` provide the RoboDojo target configuration and normalization statistics. Base model assets are downloaded separately.
|
| 6 |
+
|
| 7 |
+
Verify the downloaded file from this directory with `sha256sum -c SHA256SUMS`.
|
pretrain/SHA256SUMS
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
45dfbabba964e7f5276c738a0e5f3ebd797102a9cc57a62d3916547dfe3bb28c model.pt
|
pretrain/manifest.json
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"format": "liber0-pretrain-v1",
|
| 3 |
+
"sha256": {
|
| 4 |
+
"model.pt": "45dfbabba964e7f5276c738a0e5f3ebd797102a9cc57a62d3916547dfe3bb28c"
|
| 5 |
+
},
|
| 6 |
+
"bytes": {
|
| 7 |
+
"model.pt": 17627660073
|
| 8 |
+
}
|
| 9 |
+
}
|
pretrain/model.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:45dfbabba964e7f5276c738a0e5f3ebd797102a9cc57a62d3916547dfe3bb28c
|
| 3 |
+
size 17627660073
|