The dataset viewer is not available for this subset.
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 78, 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 54, in _get_pipeline_from_tar
current_example[field_name] = cls.DECODERS[data_extension](current_example[field_name])
~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/webdataset/webdataset.py", line 332, in torch_loads
return torch.load(io.BytesIO(data), weights_only=True)
~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/torch/serialization.py", line 1609, in load
raise pickle.UnpicklingError(_get_wo_message(str(e))) from None
_pickle.UnpicklingError: Weights only load failed. This file can still be loaded, to do so you have two options, [1mdo those steps only if you trust the source of the checkpoint[0m.
(1) In PyTorch 2.6, we changed the default value of the `weights_only` argument in `torch.load` from `False` to `True`. Re-running `torch.load` with `weights_only` set to `False` will likely succeed, but it can result in arbitrary code execution. Do it only if you got the file from a trusted source.
(2) Alternatively, to load with `weights_only=True` please check the recommended steps in the following error message.
WeightsUnpickler error: Unsupported global: GLOBAL numpy.core.multiarray._reconstruct was not an allowed global by default. Please use `torch.serialization.add_safe_globals([numpy.core.multiarray._reconstruct])` or the `torch.serialization.safe_globals([numpy.core.multiarray._reconstruct])` context manager to allowlist this global if you trust this class/function.
Check the documentation of torch.load to learn more about types accepted by default with weights_only https://pytorch.org/docs/stable/generated/torch.load.html.
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 68, 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.
LingBot-VA BEHAVIOR-1K generated training artifacts
Precomputed Wan2.2 VAE camera latents, UMT5 text embeddings, and generated training metadata for the LingBot-VA BEHAVIOR-1K experiments. This is the dataset repository. Model checkpoints are in the separate model repository.
The original demonstrations remain in
behavior-1k/2026-challenge-demos.
This repository contains no original videos, action Parquets, model weights,
training checkpoints, or runtime caches. restore.py reconstructs the
trainer's action-only Parquets from upstream data, using the saved episode
mappings. It does not rerun VAE/text encoding or require the annotation pipeline.
| Dataset | Split | Tasks | Episodes | Windows | Camera latent files |
|---|---|---|---|---|---|
b1k_main |
train | 100 | 5,000 | 167,580 | 502,740 |
b1k_main |
val | 100 | 1,000 | 31,309 | 93,927 |
radio200 |
train | 1 | 200 | 1,477 | 4,431 |
b1k_main uses the first 50 valid episodes per task for training and ordinals
190–199 for validation. Exact episode membership is saved in the manifest and
per-task mappings. radio200 includes every original turning_on_radio episode 0–199,
including the former radio validation episodes. Radio200 has no held-out
radio split and must not be treated as disjoint from the main validation set.
Contents
Archives are grouped by task and split to make the roughly 600,000 generated files practical to transfer. They are uncompressed tar files; their contents retain the exact relative paths expected by the trainer.
manifest.json # counts, original episode IDs, archive hashes
SHA256SUMS # archive and empty-embedding checksums
restore.py # verify, unpack, reconstruct action tables
b1k_main/
empty_emb.pt
train/task-0000.tar ... task-0099.tar
val/task-0000.tar ... task-0099.tar
radio200/
empty_emb.pt
train/task-0000.tar
Each archive expands to <split>/task-XXXX/meta/ and
<split>/task-XXXX/latents/. Metadata includes atomic instructions, window
boundaries, original subtask ends for referee labels, per-episode action
statistics, and local-to-original episode mappings.
Latents are normalized VAE posterior means, stored separately for the head, left wrist, and right wrist cameras. Original demonstrations are 30 Hz; RGB sampling stride is 4 and one latent interval corresponds to 16 actions. A 128-action window has 9 latent frames, including the initial conditioning frame.
The main dataset uses latents/text_bank.pt for subtask, paraphrase, and
full-task embeddings. Radio200 uses latents/text_bank_task_subtask.pt for
combined prompts, such as:
Task: Turn on the radio receiver that's on the table in the living room.
Subtask: go to the radio
Download and restore on another training server
Use Python 3.11 or newer with huggingface_hub, numpy, and pyarrow installed
(the LingBot-VA training environment already includes them). This dataset is
public and can be downloaded without authentication. For reproducible reuse, add
--revision <dataset-commit> to hf download.
hf download AkashKarnatak/lingbot-va-b1k --repo-type dataset \
--local-dir "$HOME/dataset/b1k-generated-download" \
--include 'README.md' 'manifest.json' 'SHA256SUMS' 'restore.py' 'b1k_main/**'
python "$HOME/dataset/b1k-generated-download/restore.py" \
--dataset b1k_main \
--output-root "$HOME/dataset/behavior-1k/lingbot-va-b1k" \
--demos-root "$HOME/dataset/behavior-1k/2026-challenge-demos" \
--download-missing
export B1K_DERIVED_ROOT="$HOME/dataset/behavior-1k/lingbot-va-b1k"
--download-missing fetches only upstream episode metadata and action shards
needed for the selected episodes, pinned to the upstream revision in the
manifest. Existing action shards are reused. Videos are not downloaded. Omit
that flag for an offline restore when those upstream files are already present.
Use an empty output directory for the initial restore. Archives and extracted
latents together need about 320 GB, plus space for upstream action shards and
the reconstructed action tables. The full release needs about 600,000 inodes.
For only the radio experiment:
hf download AkashKarnatak/lingbot-va-b1k --repo-type dataset \
--local-dir "$HOME/dataset/b1k-generated-download" \
--include 'README.md' 'manifest.json' 'SHA256SUMS' 'restore.py' 'radio200/**'
python "$HOME/dataset/b1k-generated-download/restore.py" \
--dataset radio200 \
--output-root "$HOME/dataset/behavior-1k/radio200_task_subtask_20260909" \
--demos-root "$HOME/dataset/behavior-1k/2026-challenge-demos" \
--download-missing
export B1K_DERIVED_ROOT="$HOME/dataset/behavior-1k/radio200_task_subtask_20260909"
Use the b1k_radio_train training configuration for radio200; it requires
instruction_mode=task_subtask and cfg_prob=hl_text_prob=0.
The main dataset uses b1k_train with the original subtask prompt contract.
Training still needs the LingBot-VA code/configuration and base model from
robbyant/lingbot-va-base, or an appropriate initialization checkpoint.
To restore selected task archives, restrict the download patterns and pass
--tasks task-0000 task-0001 and/or --splits train to restore.py. Always
download the dataset's empty_emb.pt and the root manifest.json too.
--artifacts-only skips action reconstruction; that output needs action tables
before it can train. restore.py verifies each selected archive's SHA-256,
member count, and payload size before reporting it restored.
Provenance and verification
Source artifacts were generated on the training server and frozen for this
release on 2026-09-10. The release audit verifies that every metadata window
has exactly the three expected camera files. manifest.json records archive
checksums, counts, original episode membership, and the upstream revision used
for restoration. verification.json, when present, records completed Hub
checksum verification and restoration checks.
Upstream demonstrations are MIT-licensed; their copyright notice is retained
in UPSTREAM_LICENSE. Base model and text-encoder usage remains subject to
the terms of their respective sources. No model weights are redistributed here.
- Downloads last month
- 15