LucaFrat/groot-bs16

GR00T fine-tune

Fine-tuned from nvidia/GR00T-N1.7-3B on 2026-05-27T10:45:19+00:00. Embodiment: UNITREE_G1_SONIC. Trained on 4 GPUs.

Hyperparameters

timestamp           = 2026-05-27T10:29:39+00:00
log_file            = /home/ubuntu/groot-files/logs/train-2026-05-27-102939.log
base_model          = nvidia/GR00T-N1.7-3B
embodiment_tag      = UNITREE_G1_SONIC
dataset_dir         = /home/ubuntu/groot-files/dataset
hf_dataset          = LucaFrat/dataset_100
checkpoint_dir      = /home/ubuntu/groot-files/checkpoints/g1_finetune-20260527-102938
num_gpus            = 4
visible_gpu         = <all>
global_batch_size   = 16
max_steps           = 4000
save_steps          = 500
save_total_limit    = 10
learning_rate       = 1e-4
warmup_ratio        = 0.05
weight_decay        = 1e-5
dataloader_workers  = 6
resume              = <none>
use_wandb           = 1
wandb_project       = groot-finetune
wandb_run_name      = <auto>

Checkpoints

Intermediate checkpoints (checkpoint-N/) are pushed incrementally during training; the top-level files in this repo are the final model state.