Update model card for the batch-32 step-20000 retrain
Browse files
README.md
CHANGED
|
@@ -6,29 +6,36 @@ tags:
|
|
| 6 |
- vla
|
| 7 |
datasets:
|
| 8 |
- armnet/busybox_multitask
|
|
|
|
| 9 |
---
|
| 10 |
|
| 11 |
# molmoact2_busybox_multitask
|
| 12 |
|
| 13 |
-
Full-component fine-tune of [MolmoAct2](https://huggingface.co/allenai/MolmoAct2) on SO-101 BusyBox multitask. Token embeddings stay frozen.
|
| 14 |
|
| 15 |
| | |
|
| 16 |
|---|---|
|
| 17 |
| **Init** | `allenai/MolmoAct2` (not the SO100_101 checkpoint) |
|
| 18 |
-
| **Dataset** | [armnet/busybox_multitask](https://huggingface.co/datasets/armnet/busybox_multitask) (132 episodes, 27 tasks) |
|
| 19 |
| **Cameras** | `top`, `wrist`, `front` |
|
| 20 |
-
| **Training** |
|
| 21 |
-
| **Result** | loss 0.
|
| 22 |
-
| **W&B** | [
|
| 23 |
|
| 24 |
## Checkpoints
|
| 25 |
|
| 26 |
-
`checkpoint-20000/` is the final step.
|
| 27 |
-
|
| 28 |
-
`checkpoint-5000/` · `checkpoint-10000/` · `checkpoint-15000/` · `checkpoint-20000/`
|
| 29 |
|
| 30 |
```python
|
| 31 |
from huggingface_hub import snapshot_download
|
| 32 |
|
| 33 |
-
ckpt = snapshot_download(
|
|
|
|
|
|
|
|
|
|
|
|
|
| 34 |
```
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 6 |
- vla
|
| 7 |
datasets:
|
| 8 |
- armnet/busybox_multitask
|
| 9 |
+
base_model: allenai/MolmoAct2
|
| 10 |
---
|
| 11 |
|
| 12 |
# molmoact2_busybox_multitask
|
| 13 |
|
| 14 |
+
Full-component fine-tune of [MolmoAct2](https://huggingface.co/allenai/MolmoAct2) on SO-101 BusyBox multitask. Token embeddings stay frozen. This card is the 4× H200 retrain at effective batch 32.
|
| 15 |
|
| 16 |
| | |
|
| 17 |
|---|---|
|
| 18 |
| **Init** | `allenai/MolmoAct2` (not the SO100_101 checkpoint) |
|
| 19 |
+
| **Dataset** | [armnet/busybox_multitask](https://huggingface.co/datasets/armnet/busybox_multitask) (132 episodes / 25,734 frames, 27 tasks) |
|
| 20 |
| **Cameras** | `top`, `wrist`, `front` |
|
| 21 |
+
| **Training** | 4× H200, effective batch 32 (8/GPU), 20k steps, horizon 30, gradient checkpointing off, bf16 |
|
| 22 |
+
| **Result** | loss 0.220 → 0.005. 14 h 23 m. Finished 2026-10-27 12:00 UTC. 5.046B learnable / 5.442B total |
|
| 23 |
+
| **W&B** | [dhseedj1](https://wandb.ai/pravsels/molmoact2_busybox_multitask/runs/dhseedj1) |
|
| 24 |
|
| 25 |
## Checkpoints
|
| 26 |
|
| 27 |
+
`checkpoint-20000/` is the final step. Tags `step-5000`, `step-10000`, `step-15000`, and `step-20000` point at this run.
|
|
|
|
|
|
|
| 28 |
|
| 29 |
```python
|
| 30 |
from huggingface_hub import snapshot_download
|
| 31 |
|
| 32 |
+
ckpt = snapshot_download(
|
| 33 |
+
"armnet/molmoact2_busybox_multitask",
|
| 34 |
+
revision="step-20000",
|
| 35 |
+
allow_patterns="checkpoint-20000/**",
|
| 36 |
+
)
|
| 37 |
```
|
| 38 |
+
|
| 39 |
+
## Previous run
|
| 40 |
+
|
| 41 |
+
The 2026-10-25 1× H200 batch-4 fine-tune ([olalqzcn](https://wandb.ai/pravsels/molmoact2_busybox_multitask/runs/olalqzcn)) is still on this repo at `olalqzcn-step-5000`, `olalqzcn-step-10000`, `olalqzcn-step-15000`, and `olalqzcn-step-20000`. Load that checkpoint with `revision="olalqzcn-step-20000"`.
|