Downloads · 30 days
10
24% of all-time downloads
xpuenabler/MolmoAct2-LIBERO-ViT10L
MolmoAct2-LIBERO-ViT10L is a robotics model from xpuenabler. Use it for the robotics task on the model card, and read the license before you ship it in a product. The card lists the license as apache-2.0.
allenai/MolmoAct2-LIBERO의 Vision Encoder를 25층 → 10층(2.5×) 으로 압축한 VLA. LLM(Qwen3-4B 36L)·Action Expert는 원본과 동일. Standalone HF 체크포인트 (trustremotecode).
Downloads · 30 days
10
24% of all-time downloads
All-time downloads
42
Public
Parameters
5.2B
10.4 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors10.4 GB · 100%
How the weights are stored.
BF165.2B · 100%
From the Hugging Face model README
allenai/MolmoAct2-LIBERO의 Vision Encoder를 25층 → 10층(2.5×) 으로 압축한 VLA.
LLM(Qwen3-4B 36L)·Action Expert는 원본과 동일. Standalone HF 체크포인트 (trust_remote_code).
| spatial | object | goal | long | 전체 | |
|---|---|---|---|---|---|
| 본 모델 (ViT 10L) | 100 | 100 | 94 | 90 | 96.0% |
| 원본 teacher (25L) | 98 | 100 | 100 | 98 | 99.0% |
[0,1,2,3,5,6,7,8,9,24]만 유지, 가중치 상속. connector taps vit_layers=[-1,-4].핵심 발견: embed-KD 생략 시 71.5%로 붕괴(+24%p 차이). 최신 선택기법(탐색/병합)은 강한 healing 하에서 이득 없음.
from transformers import AutoModelForImageTextToText, AutoProcessor
model = AutoModelForImageTextToText.from_pretrained("xpuenabler/MolmoAct2-LIBERO-ViT10L", trust_remote_code=True, torch_dtype="bfloat16")
proc = AutoProcessor.from_pretrained("xpuenabler/MolmoAct2-LIBERO-ViT10L", trust_remote_code=True)
out = model.predict_action(processor=proc, images=[front, wrist], task="pick up the bowl",
state=state, norm_tag="libero", inference_action_mode="continuous")
lerobot 학습/평가: --policy.type=molmoact2 --policy.checkpoint_path=<this repo>.
training/ 폴더에 재현 팩 포함: 단계별 가이드(TRAINING.md), 전체 실험 기록(EXPERIMENTS.md),
keep-set/FFN 채널 확정 JSON, embed-KD·조립·패키징 스크립트. training/README.md부터 참조.