MolmoAct2 YAM Touch/Grab LoRA
Private fine-tuned checkpoint for the YAM touch/grab multitask dataset.
Source
- Base checkpoint:
allenai/MolmoAct2-LIBERO
- Training run:
20260602_212941
- Best checkpoint:
molmoact2_lora_20260602_212941/checkpoints/000779
- Training split:
andon/yam_touch_grab_201_10hz_train
- Eval split:
andon/yam_touch_grab_201_10hz_eval
Fine-tuning config
- VLM LoRA enabled:
true
- Action-expert LoRA enabled:
false
- Rank:
16
- Alpha:
32
- Dropout:
0.05
- Bias:
none
- Batch size:
16
- Precision:
bfloat16
- Chunk size:
10
- Action steps:
10
- Freeze embedding:
true
- Train action expert only:
false
Result
- Best validation step:
779
- Best validation epoch:
1.0
- Best validation loss:
0.0856607579626143
- Training stopped cleanly by early stopping at epoch
6.0.
The model files are uploaded at the repository root from the checkpoint's pretrained_model/ directory.
