num_frames int64 | token_dim int64 | horizon int64 | action_dim int64 | num_samples int64 | num_heldout int64 | stride int64 | discount float64 | dtype string | config string | checkpoint string | repo_id string | reward_scheme string | terminal_success bool | value_support list | has_borrowed_return bool |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
279,534 | 2,048 | 16 | 12 | 16 | 8 | 1 | 0.99 | float32 | pi05_robocasa_PrepareCoffee_rlt | /data5/jellyho/ACRFT/openpi/checkpoints/pi05_robocasa_PrepareCoffee_rlt/PrepareCoffee_rlt5_pardec_noprop/70000 | jellyho/robocasa365-PrepareCoffee | sparse | true | [
0,
1
] | true |
YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
acrft-annot-noprop
RLT annotation for AC-RFT critic training: raw memmaps (.dat) + meta.json, read directly by
scripts/train_rlt_critic.py and scripts/eval_rlt_critic.py in the openpi fork.
Shapes are in meta.json: rl_token [T, D], base_action [T, N, H, A], action_chunk [T, H, A], reward/mc_return/done/episode_index/frame_index [T], base_action_heldout [T, num_heldout, H, A]. dtype and reward_scheme are in meta.json. Load with numpy.memmap.
from huggingface_hub import snapshot_download
d = snapshot_download("jellyho/acrft-annot-noprop", repo_type="dataset")
uv run scripts/train_rlt_critic.py --data $d --kind arq --steps 100000 ...
- Downloads last month
- 32