Modelpublic

RobotLearningVLA/eval3-smolvla-v4-toy-merged-50k-unfrozen

sourceHugging Faceapache-2.0updated 4mo agoView on Hugging Face

0likes6downloads

Model Card

eval3-smolvla-v4-toy-merged-50k-unfrozen

SmolVLA policy continued from RobotLearningVLA/eval3-smolvla-v4-toy-merged-25k-unfrozen on RobotLearningVLA/dataset_v4_toy_merged for Eval3 (SO-101 toy task, merged v4 dataset).

Training

Field	Value
Base checkpoint	`RobotLearningVLA/eval3-smolvla-v4-toy-merged-25k-unfrozen` (local: `outputs/train/smolvla_v4_toy_merged/checkpoints/025000/pretrained_model`)
Continuation steps	25,000 additional (checkpoint `025000` of continue run)
Total effective fine-tuning	50,000 steps (25k initial + 25k continue)
Final train loss	~0.048 (step 25000, continue run log)
Vision / expert	Unfrozen (`train_expert_only=false`, `freeze_vision_encoder=false`)
Brev run	`radical-salmon-hoverfly` (H100)

Eval3 dataset / policy config

—Dataset: RobotLearningVLA/dataset_v4_toy_merged
—rename_map: observation.images.front → observation.images.camera1
—empty_cameras: 2 (pad to SmolVLA expected camera count)

Local continue-run checkpoints also saved at steps 5k–25k on Brev under outputs/train/smolvla_v4_toy_merged_50k/; this Hub repo is the 50k (025000 continue) pretrained_model only.

Usage

bash

lerobot-record ... --policy.path=RobotLearningVLA/eval3-smolvla-v4-toy-merged-50k-unfrozen

Eval status

—Gate matrix / robot eval: TODO (see Eval3 protocol in repo v4_qa_output/).