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/python3.12/site-packages/datasets/inspect.py", line 286, in get_dataset_config_info
                  for split_generator in builder._split_generators(
                                         ^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.12/site-packages/datasets/packaged_modules/webdataset/webdataset.py", line 80, in _split_generators
                  first_examples = list(islice(pipeline, self.NUM_EXAMPLES_FOR_FEATURES_INFERENCE))
                                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.12/site-packages/datasets/packaged_modules/webdataset/webdataset.py", line 33, in _get_pipeline_from_tar
                  for filename, f in tar_iterator:
                                     ^^^^^^^^^^^^
                File "/usr/local/lib/python3.12/site-packages/datasets/utils/track.py", line 49, in __iter__
                  for x in self.generator(*self.args):
                           ^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.12/site-packages/datasets/utils/file_utils.py", line 1380, in _iter_from_urlpath
                  yield from cls._iter_tar(f)
                File "/usr/local/lib/python3.12/site-packages/datasets/utils/file_utils.py", line 1331, in _iter_tar
                  stream = tarfile.open(fileobj=f, mode="r|*")
                           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.12/tarfile.py", line 1886, in open
                  t = cls(name, filemode, stream, **kwargs)
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.12/tarfile.py", line 1762, in __init__
                  self.firstmember = self.next()
                                     ^^^^^^^^^^^
                File "/usr/local/lib/python3.12/tarfile.py", line 2756, in next
                  raise ReadError(str(e)) from None
              tarfile.ReadError: truncated header
              
              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 65, in compute_split_names_from_streaming_response
                  for split in get_dataset_split_names(
                               ^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.12/site-packages/datasets/inspect.py", line 340, in get_dataset_split_names
                  info = get_dataset_config_info(
                         ^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.12/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.

eval3_track3_vl_pairs_v2

ObjectVLA-style vision-language co-training data for the Track 3 TOY 3-celeb Eval 3 task. v2 fixes two bugs in v1:

  1. Oriented quad instead of axis-aligned xyxy. Tilted portraits no longer bloat the box with non-portrait area.
  2. ArcFace-verified labels. Each portrait's label is confirmed by ArcFace against precomputed Swift/Obama/LeCun prototypes; mismatches with the layout-string derivation are recorded and corrected.

Companion to HBOrtiz/so101_eval3_track3_v3_baseline.

Contents

  • manifest.parquet — VL pair rows (one per portrait per caption type).
  • data.tar.zst — compressed archive with images/chunk-*/ wrist-cam JPEGs and references/*.jpg reference photos. Extract with tar --zstd -xf data.tar.zst.
  • _stats.json — dataset statistics.

Schema (new columns vs v1)

column type meaning
quad_corners_norm list[4][2] NEW — 4 (x, y) corners of the printed paper in wrist-cam coords (normalized). The primary geometric ground truth for KLAL polygon targets.
bbox_xyxy_norm list[4] Axis-aligned bbox derived from the quad — for PaliGemma <loc> token compat. Loose for tilted portraits.
celeb_name str PRIMARY identity label (ArcFace-verified, falls back to derived if face not detected)
derived_celeb_name str NEW — re-derived label from layout-string logic (kept for analysis)
verification_ok bool NEW — True iff ArcFace agreed with derived (and score ≥ 0.30)
verification_score float NEW — ArcFace cosine to best-match prototype

Other columns same as v1 (image_path, reference_image_path, prompt, target, bbox_refit_ok, celeb_slug, caption_type, episode, episode_index, frame_idx, pid).

Bbox geometry: oriented quad vs axis-aligned bbox

For a tilted printed portrait (40° rotation example), the axis-aligned xyxy box bloats by ~40-50%. The 4-corner quad describes the actual rotated rectangle.

For KLAL attention-target maps, build them from quad_corners_norm (use as a polygon mask). For PaliGemma <loc> token prompts that need axis-aligned coords, use bbox_xyxy_norm.

ArcFace verification

For each portrait, the script perspective-warps the 4-corner quad to a 320×320 crop, runs InsightFace + ArcFace, and matches to precomputed Swift/Obama/LeCun prototypes (10 photos per celeb from the scraped bank, L2-normalized mean embedding).

Three outcomes per portrait:

  • verification_ok=True: ArcFace best-match equals derived AND score ≥ 0.30. Label trusted.
  • verification_ok=False, verification_score > 0: ArcFace disagreed with derivation. celeb_name uses ArcFace's verdict.
  • verification_score=0: No face detected in crop. celeb_name falls back to derived.

Filter at training time with df.query("verification_ok") if you want only high-confidence pairs.

Stats

See _stats.json.

Downloads last month
5