Datasets:

---

t1an

/

ACWM-Phys

like 0

Hi, thank you for releasing ACWM-Phys.

We found a possible incomplete upload in:

kinematics/reacher/ind_train

At dataset revision:

4017e5f5900a7a8590d4e944b6a84f55df75a0c1

we observed the following:

• metadata.pt contains 40,000 entries.
• The referenced video filenames uniquely cover episode_0.mp4 through
  episode_39999.mp4, with no missing indices in the metadata.
• A complete traversal of the Hugging Face repository tree returns only:
  • 9,993 episode_*.mp4 files
  • metadata.pt
  • metadata_part_0.pt through metadata_part_4.pt
• Therefore, 30,007 MP4 files referenced by metadata.pt are not available
  in the current Hub repository.

We also compared a complete local download against the Hugging Face repository
tree. The two filename sets are identical, so this does not appear to be an
interrupted local download.

Minimal reproduction:

from huggingface_hub import HfApi

api = HfApi()
items = list(
    api.list_repo_tree(
        repo_id="t1an/ACWM-Phys",
        repo_type="dataset",
        revision="4017e5f5900a7a8590d4e944b6a84f55df75a0c1",
        path_in_repo="kinematics/reacher/ind_train",
        recursive=True,
    )
)

mp4_files = [
    item.path
    for item in items
    if item.path.endswith(".mp4")
]

print(len(items))      # 9999
print(len(mp4_files))  # 9993

There are also two related inconsistencies:
The dataset README says that each environment contains 1,000 training
trajectories.
hf_upload.log reports that the source
kinematics/reacher/ind_train directory contained 40,011 files.
Could you please clarify the intended released training-set size for Reacher?
If 40,000 trajectories are intended to be public, could you upload the
remaining 30,007 MP4 files? If 9,993 trajectories are the intended public
subset, could you provide an updated metadata.pt or an exact manifest
containing only the released videos?
Thanks!