Dataset Viewer
Duplicate
The dataset viewer is not available for this split.
Cannot load the dataset split (in streaming mode) to extract the first rows.
Error code:   StreamingRowsError
Exception:    TypeError
Message:      Couldn't cast array of type
struct<a1: string, a2: string, a7: string, a8: string, b1: string, b2: string, b7: string, b8: string, c2: string, c7: string, c8: string, d7: string, d8: string, e1: string, e2: string, e7: string, f1: string, f3: string, f5: string, f7: string, g1: string, g2: string, g5: string, g8: string, h1: string, h2: string, h5: string, h8: string>
to
{'a2': Value('string'), 'a7': Value('string'), 'a8': Value('string'), 'b1': Value('string'), 'b2': Value('string'), 'b7': Value('string'), 'b8': Value('string'), 'c1': Value('string'), 'c2': Value('string'), 'c3': Value('string'), 'c7': Value('string'), 'c8': Value('string'), 'd4': Value('string'), 'd7': Value('string'), 'd8': Value('string'), 'e2': Value('string'), 'e7': Value('string'), 'e8': Value('string'), 'f1': Value('string'), 'f2': Value('string'), 'f6': Value('string'), 'f8': Value('string'), 'g1': Value('string'), 'g7': Value('string')}
Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/utils.py", line 147, 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 127, 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 483, 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 2840, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2373, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2398, 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 536, in _iter_arrow
                  for key, pa_table in iterator:
                                       ^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 419, 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 343, in _generate_tables
                  self._cast_table(pa_table, json_field_paths=json_field_paths),
                  ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
                  pa_table = table_cast(pa_table, self.info.features.arrow_schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2378, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2312, in cast_table_to_schema
                  cast_array_to_feature(
                  ~~~~~~~~~~~~~~~~~~~~~^
                      table[name] if name in table_column_names else pa.array([None] * len(table), type=schema.field(name).type),
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                      feature,
                      ^^^^^^^^
                  )
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 1861, in wrapper
                  return pa.chunked_array([func(chunk, *args, **kwargs) for chunk in array.chunks])
                                           ~~~~^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2158, in cast_array_to_feature
                  raise TypeError(f"Couldn't cast array of type\n{_short_str(array.type)}\nto\n{_short_str(feature)}")
              TypeError: Couldn't cast array of type
              struct<a1: string, a2: string, a7: string, a8: string, b1: string, b2: string, b7: string, b8: string, c2: string, c7: string, c8: string, d7: string, d8: string, e1: string, e2: string, e7: string, f1: string, f3: string, f5: string, f7: string, g1: string, g2: string, g5: string, g8: string, h1: string, h2: string, h5: string, h8: string>
              to
              {'a2': Value('string'), 'a7': Value('string'), 'a8': Value('string'), 'b1': Value('string'), 'b2': Value('string'), 'b7': Value('string'), 'b8': Value('string'), 'c1': Value('string'), 'c2': Value('string'), 'c3': Value('string'), 'c7': Value('string'), 'c8': Value('string'), 'd4': Value('string'), 'd7': Value('string'), 'd8': Value('string'), 'e2': Value('string'), 'e7': Value('string'), 'e8': Value('string'), 'f1': Value('string'), 'f2': Value('string'), 'f6': Value('string'), 'f8': Value('string'), 'g1': Value('string'), 'g7': Value('string')}

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.

SightRead — Chess pack v0.2 (ground truth + prompt)

Procedurally generated chess positions for the SightRead benchmark: exact position extraction from rendered board images by vision-language models.

Not whether a model can play chess — whether it can see the board.

What's in here

File Purpose
pack.yaml field spec: 10 fields with comparator types + grounding trap
prompt.md the extraction prompt shown to every model
gt/*.json 6 playout positions (3 busy 20-28 pieces, 3 sparse 8-14)

Each GT state contains: side to move, the full square→piece map, both king squares, per-side piece and pawn counts, the rank-1 piece multiset, and the a-file occupancy count, plus passthrough provenance keys (fen, seed, image, generator).

Reachable and contamination-resistant by construction

Positions are sampled from seeded random chess playouts that start from the standard starting position and play legal moves (captures preferred) until the piece count enters the target band. Every position therefore:

  • arose from a real legal game — pawn structures, castling rights and promotions are all genuine; no "technically legal but impossible-looking" boards (e.g. same-color double bishops can only appear if a promotion in the actual playout produced them);
  • matches no position in FEN databases or training corpora (seeded random play, not openings theory).

Images are re-generable byte-identically (same python-chess major version — the render extra pins it):

python tools/render_chess.py --out packs/chess --count 6 --seed 20260828

Rendering is PIL-only (Unicode piece glyphs from Segoe UI Symbol), so the pack has no cairo/svg dependencies. Images live in packs/chess/images/ (gitignored by design — they are re-renderable from the committed seeds).

Provenance: v0.1 placed random piece sets on an empty board with static legality checks; a reviewer caught a same-color double-bishop position there — legal only via promotion, but impossible-looking — and v0.2 replaced that generator with playout sampling.

Grounding trap

prompt.md embeds a full extracted answer for the standard starting position as a format example, which matches NO image (generated positions have 8-28 pieces). A model whose extracted board equals it echoed its memorized starting position instead of reading the image — SightRead's grounding-fidelity signal, declared in pack.yaml and flagged automatically by the scorer.

Scoring

Format-agnostic: responses are parsed line-wise (keys matched case/dash/space-insensitively) with a fallback pass that derives counts, king squares, rank-1 pieces and a-file occupancy from the parsed board map when possible. Boards score as exact mappings; rank-1 pieces as multisets; numbers exact; squares case-insensitive text. Full logic in src/sightread/ at the repository above.

Citation

@software{sightread2026,
  title  = {SightRead: Exact Game-State Extraction Benchmark for Vision-Language Models},
  author = {AnsteinHuynh},
  year   = {2026},
  url    = {https://github.com/AnsteinHuynh/sightread-bench}
}
Downloads last month
39