Dataset Viewer
The dataset viewer is not available for this subset.
Cannot get the split names for the config 'default' of the dataset.
Exception:    SplitsNotFoundError
Message:      The split names could not be parsed from the dataset config.
Traceback:    Traceback (most recent call last):
                File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 286, in get_dataset_config_info
                  for split_generator in builder._split_generators(
                                         ~~~~~~~~~~~~~~~~~~~~~~~~~^
                      StreamingDownloadManager(base_path=builder.base_path, download_config=download_config)
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  )
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/webdataset/webdataset.py", line 81, in _split_generators
                  first_examples = list(islice(pipeline, self.NUM_EXAMPLES_FOR_FEATURES_INFERENCE))
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/webdataset/webdataset.py", line 32, in _get_pipeline_from_tar
                  fs: fsspec.AbstractFileSystem = fsspec.filesystem("memory")
                                                  ~~~~~~~~~~~~~~~~~^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/fsspec/registry.py", line 302, in filesystem
                  cls = get_filesystem_class(protocol)
                File "/usr/local/lib/python3.14/site-packages/fsspec/registry.py", line 239, in get_filesystem_class
                  raise ValueError(f"Protocol not known: {protocol}")
              ValueError: Protocol not known: memory
              
              The above exception was the direct cause of the following exception:
              
              Traceback (most recent call last):
                File "/src/services/worker/src/worker/job_runners/config/split_names.py", line 71, in compute_split_names_from_streaming_response
                  for split in get_dataset_split_names(
                               ~~~~~~~~~~~~~~~~~~~~~~~^
                      path=dataset,
                      ^^^^^^^^^^^^^
                      config_name=config,
                      ^^^^^^^^^^^^^^^^^^^
                      token=hf_token,
                      ^^^^^^^^^^^^^^^
                  )
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 340, in get_dataset_split_names
                  info = get_dataset_config_info(
                      path,
                  ...<6 lines>...
                      **config_kwargs,
                  )
                File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 291, in get_dataset_config_info
                  raise SplitsNotFoundError("The split names could not be parsed from the dataset config.") from err
              datasets.inspect.SplitsNotFoundError: The split names could not be parsed from the dataset config.

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

h200-nav-eval-exchange

Cross-checking two boxes' open-loop evaluation of the pooled move to run. Everything here is raw physical units (m/s, rad/s), full 32-step chunk, denoise 10, seed 0.

h200_nav_preds_d1.npz (H200 box, 1 draw)

targets                      (512, 32, 23)
pred::init (pre-SFT)
pred::SFT-50k (LoRA)
pred::nav (pooled)           step-30000
frame::dataset_index         index into the eval dataset (0..107,024)
frame::source_episode_index  original episode, 0..199
frame::source_start_frame    segment start in that episode
frame::frame_in_episode      offset inside the segment
# absolute source frame = source_start_frame + frame_in_episode

Frame ids were reconstructed from the seed (numpy rng.choice, not a DataLoader shuffle) and then VERIFIED by re-reading the dataset at those indices: 9/9 sampled rows match the stored targets to 1.2e-08. The other box's seed reconstruction failed and was redone by fingerprinting; ours reproduces because the picks are drawn directly in main().

Scored with pi05_b1k_pytorch_local + pi05-b1k-sft50k-merged/assets for init and SFT-50k (turning_on_radio z-score) and pi05_b1k_nav + its own assets (quantile) for nav, then converted back to physical units per model — R² over a GROUP of axes is not affine-invariant, so physical units are the only space where the columns mean the same thing.

evalsets/ — per-goal eval sets, built from the 16-task SUBSET manifest

move_to_nav_fridge_v3.0   400 segments / 2 tasks  / 290,892 frames
move_to_nav_floors_v3.0   896 segments / 4 tasks  / 387,485 frames

meta/ and data/ only. Video is symlinked into the H200's nav-subset tree, so relink against your own download: ln -s <2026-challenge-demos>/videos <dataset>/videos. The video pointers in meta/episodes are the original release's chunk/file/timestamp values. depth features are already dropped — only RGB was fetched on that box.

floors is the strongest cross-task pooling condition available in the subset: the same goal in scrubbing_bathroom_floor, sorting_bottles_cans_and_paper, unloading_the_car and vacuuming_floors.

Downloads last month
54