The dataset viewer is not available for this split.
Error code: StreamingRowsError
Exception: ValueError
Message: Expected object or value
Traceback: Traceback (most recent call last):
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 291, in _generate_tables
pa_table = paj.read_json(
io.BytesIO(batch), read_options=paj.ReadOptions(block_size=block_size)
)
File "pyarrow/_json.pyx", line 364, in pyarrow._json.read_json
File "pyarrow/error.pxi", line 155, in pyarrow.lib.pyarrow_internal_check_status
File "pyarrow/error.pxi", line 92, in pyarrow.lib.check_status
raise convert_status(status)
pyarrow.lib.ArrowInvalid: JSON parse error: Column() changed from object to string in row 0
During handling of the above exception, another exception occurred:
Traceback (most recent call last):
File "/src/services/worker/src/worker/utils.py", line 149, in get_rows_or_raise
return get_rows(
dataset=dataset,
...<4 lines>...
column_names=column_names,
)
File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
return func(*args, **kwargs)
File "/src/services/worker/src/worker/utils.py", line 129, in get_rows
rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
File "/src/services/worker/src/worker/utils.py", line 544, in safe_iter
yield from ds.decode(False) if ds.features else ds
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 3052, in __iter__
for key, example in ex_iterable:
^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2559, in __iter__
for key, pa_table in self._iter_arrow():
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2584, in _iter_arrow
for key, pa_table in self.ex_iterable._iter_arrow():
~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 558, in _iter_arrow
for key, pa_table in iterator:
^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 440, in _iter_arrow
for key, pa_table in self.generate_tables_fn(**gen_kwags):
~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 306, in _generate_tables
batch = json_encode_fields_in_json_lines(original_batch, json_field_paths)
File "/usr/local/lib/python3.14/site-packages/datasets/utils/json.py", line 111, in json_encode_fields_in_json_lines
examples = [ujson_loads(line) for line in original_batch.splitlines()]
~~~~~~~~~~~^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/utils/json.py", line 20, in ujson_loads
return pd.io.json.ujson_loads(*args, **kwargs)
~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^
ValueError: Expected object or valueNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
egoverse-v10-randref
A random same-task reference control for y6hQ Q2 ("randomly selected same-task references under matched training settings").
manifest/pairs_randref.npzstarts from WATCH's training manifest,egoverse-v10-state/manifest/pairs.npz. That is the file the experimentcosmos_predict2p5_2b_egoverse_ret_v3_448_statereads;pairs_clean.npzbelongs to the separate_cleanexperiment.- Every reference slot is replaced by a uniform draw from the windows with the same (source, fine task class) as the original reference. Per query, the number of references and their source/class composition are identical to WATCH's: 10.52 refs per query, 60.4% cross-source, 73.2% same fine class. The only thing removed is the motion ranking (DTW + handedness).
- The §C.3 constraints are kept: no reference from the query's own episode (MimicDroid play excepted, non-overlapping windows only), at most 2 windows per episode per query, no duplicates.
- Held-out episodes are excluded from the pool, whereas a small share of WATCH's original references point to held-out episodes.
match_distandmatch_speedare zeroed. Training does not read them (EgoVerseRetDataset.__getitem__samples onematch_idxslot uniformly).manifest/query_eps_nomd.json: all 13,295 query episodes except the 20 MimicDroid ones (1,028,812 of 1,045,007 rows). MimicDroid video is not in the EgoVerse R2 bucket, so the from-scratch download cannot train on it; no human query references a MimicDroid window, so dropping it as a query removes it completely. Builder:responses/y6hQ/scripts/build_query_eps_nomd.py. Used by the experimentcosmos_predict2p5_2b_egoverse_ret_v3_448_state_randref_nomd_15k.- This repo holds only
manifest/pairs_randref.npz,manifest/query_eps_nomd.jsonandmanifest/pairs_randref_check.json. Every other manifest file (windows.npz, stats,query_eps_*,t5_embeddings.pkl, …) is the unchanged one fromJeongeun/egoverse-v10-state. - Checks:
manifest/pairs_randref_check.json - Builder:
responses/y6hQ/scripts/build_randref_manifest.py <out> pairs.npz(seed 0)
Use
Put pairs_randref.npz next to the v10-state manifest, either by copying it into egoverse-v10-state/manifest/ or by making a directory that symlinks the v10-state manifest files and adds this one:
hf download Jeongeun/egoverse-v10-randref --repo-type dataset --local-dir egoverse-v10-randref
cp egoverse-v10-randref/manifest/pairs_randref.npz egoverse-v10-randref/manifest/query_eps_nomd.json egoverse-v10-state/manifest/
For the full download -> 2-node training guide see docs/egoverse_randref_from_scratch.md in github.com/jeongeun980906/cosmos_policy (experiment cosmos_predict2p5_2b_egoverse_ret_v3_448_state_randref_nomd_15k: 2 nodes x 8 GPUs, global batch 32, 15K iterations, MimicDroid excluded).
To instead run the WATCH experiment with only the pairs file changed:
experiment=cosmos_predict2p5_2b_egoverse_ret_v3_448_state
dataloader_train.dataset.manifest_dir=<...>/egoverse-v10-state/manifest
dataloader_train.dataset.pairs_name=pairs_randref.npz
For a cheaper run, use the q25 arm with the same query_episode_set (query_eps_q25.json) and the same 2K iterations. Compare against WATCH q25 at the 25% point of Fig. 9.
Verification: motion agreement of one training draw per query
3,000 human queries, one reference sampled per query the way training samples it. Metrics are the Table 5 ones from 222_pairstat_dyn.agree. These numbers were measured on the earlier version built from pairs_clean.npz. The builder is unchanged, so the current version is expected to give the same result, but it has not been re-run.
| manifest | cross-source | speed corr | shape cos | |log PL| | speed >3× |
|---|---|---|---|---|---|
WATCH (pairs_clean) |
59.7% | 0.170 | 0.251 | 0.725 | 19.1% |
random-ref (from pairs_clean) |
60.0% | −0.001 | −0.002 | 0.931 | 34.3% |
- Downloads last month
- 28