The dataset viewer is not available for this split.
Error code: StreamingRowsError
Exception: CastError
Message: Couldn't cast
compared: int64
match: int64
file_bytes_match: int64
markdown_match: int64
included_native_rows: int64
excluded_native_rows: list<item: null>
child 0, item: null
parse_mismatches: list<item: null>
child 0, item: null
aggregate_files: int64
hash_definition: string
snapshot: string
benchmarks: struct<dp_bench: struct<files: int64, bytes: int64, records: struct<judge/clean/gpt-6.1-sol-medium-r (... 26980 chars omitted)
child 0, dp_bench: struct<files: int64, bytes: int64, records: struct<judge/clean/gpt-6.1-sol-medium-r0/records.jsonl: (... 2128 chars omitted)
child 0, files: int64
child 1, bytes: int64
child 2, records: struct<judge/clean/gpt-6.1-sol-medium-r0/records.jsonl: int64, judge/clean/gpt-6.1-sol-medium-r0/req (... 2083 chars omitted)
child 0, judge/clean/gpt-6.1-sol-medium-r0/records.jsonl: int64
child 1, judge/clean/gpt-6.1-sol-medium-r0/requests.jsonl: int64
child 2, judge/clean/gpt-6.1-sol-medium-r1/records.jsonl: int64
child 3, judge/clean/gpt-6.1-sol-medium-r1/requests.jsonl: int64
child 4, judge/tilt_2_4/gpt-6.1-sol-medium-r0/records.jsonl: int64
child 5, judge/tilt_2_4/gpt-6.1-sol-medium-r0/requests.jsonl: int64
child 6, native/clean/azure_layout_high_res_formulas/scores.jsonl: int64
child 7, native/clean/cohere_parse_v5/scores.jsonl: int64
child 8, native/clean/datalab_accurate/scores.jsonl: int64
child 9, native/clean/extend_performance
...
ocr_4_1/scores.jsonl: int64
child 20, native/tilt_2_4/pulse_ultra_2/scores.jsonl: int64
child 21, native/tilt_2_4/reducto_r1/scores.jsonl: int64
child 22, pages.jsonl: int64
child 23, parses/clean/azure_layout_high_res_formulas/records.jsonl: int64
child 24, parses/clean/cohere_parse_v5/records.jsonl: int64
child 25, parses/clean/datalab_accurate/records.jsonl: int64
child 26, parses/clean/extend_performance_agentic_formulas/records.jsonl: int64
child 27, parses/clean/llamaparse_agentic_plus/records.jsonl: int64
child 28, parses/clean/mistral_ocr_4_1/records.jsonl: int64
child 29, parses/clean/pulse_ultra_2/records.jsonl: int64
child 30, parses/clean/reducto_r1/records.jsonl: int64
child 31, parses/tilt_2_4/azure_layout_high_res_formulas/records.jsonl: int64
child 32, parses/tilt_2_4/cohere_parse_v5/records.jsonl: int64
child 33, parses/tilt_2_4/datalab_accurate/records.jsonl: int64
child 34, parses/tilt_2_4/extend_performance_agentic_formulas/records.jsonl: int64
child 35, parses/tilt_2_4/llamaparse_agentic_plus/records.jsonl: int64
child 36, parses/tilt_2_4/mistral_ocr_4_1/records.jsonl: int64
child 37, parses/tilt_2_4/pulse_ultra_2/records.jsonl: int64
child 38, parses/tilt_2_4/reducto_r1/records.jsonl: int64
native_validation: string
native_parse_matches: string
native_backfill: string
to
{'snapshot': Value('string'), 'benchmarks': {'dp_bench': {'files': Value('int64'), 'bytes': Value('int64'), 'records': {'judge/clean/gpt-6.1-sol-medium-r0/records.jsonl': Value('int64'), 'judge/clean/gpt-6.1-sol-medium-r0/requests.jsonl': Value('int64'), 'judge/clean/gpt-6.1-sol-medium-r1/records.jsonl': Value('int64'), 'judge/clean/gpt-6.1-sol-medium-r1/requests.jsonl': Value('int64'), 'judge/tilt_2_4/gpt-6.1-sol-medium-r0/records.jsonl': Value('int64'), 'judge/tilt_2_4/gpt-6.1-sol-medium-r0/requests.jsonl': Value('int64'), 'native/clean/azure_layout_high_res_formulas/scores.jsonl': Value('int64'), 'native/clean/cohere_parse_v5/scores.jsonl': Value('int64'), 'native/clean/datalab_accurate/scores.jsonl': Value('int64'), 'native/clean/extend_performance_agentic_formulas/scores.jsonl': Value('int64'), 'native/clean/llamaparse_agentic_plus/scores.jsonl': Value('int64'), 'native/clean/mistral_ocr_4_1/scores.jsonl': Value('int64'), 'native/clean/pulse_ultra_2/scores.jsonl': Value('int64'), 'native/clean/reducto_r1/scores.jsonl': Value('int64'), 'native/tilt_2_4/azure_layout_high_res_formulas/scores.jsonl': Value('int64'), 'native/tilt_2_4/cohere_parse_v5/scores.jsonl': Value('int64'), 'native/tilt_2_4/datalab_accurate/scores.jsonl': Value('int64'), 'native/tilt_2_4/extend_performance_agentic_formulas/scores.jsonl': Value('int64'), 'native/tilt_2_4/llamaparse_agentic_plus/scores.jsonl': Value('int64'), 'native/tilt_2_4/mistral_ocr_4_1/scores.jsonl': Value('int64'), 'native/tilt_2_4
...
4'), 'native/tilt_2_4/llamaparse_agentic_plus/scores.jsonl': Value('int64'), 'native/tilt_2_4/mistral_ocr_4_1/scores.jsonl': Value('int64'), 'native/tilt_2_4/pulse_ultra_2/scores.jsonl': Value('int64'), 'native/tilt_2_4/reducto_r1/scores.jsonl': Value('int64'), 'pages.jsonl': Value('int64'), 'parses/clean/azure_layout_high_res_formulas/records.jsonl': Value('int64'), 'parses/clean/cohere_parse_v5/records.jsonl': Value('int64'), 'parses/clean/datalab_accurate/records.jsonl': Value('int64'), 'parses/clean/extend_performance_agentic_formulas/records.jsonl': Value('int64'), 'parses/clean/llamaparse_agentic_plus/records.jsonl': Value('int64'), 'parses/clean/mistral_ocr_4_1/records.jsonl': Value('int64'), 'parses/clean/pulse_ultra_2/records.jsonl': Value('int64'), 'parses/clean/reducto_r1/records.jsonl': Value('int64'), 'parses/tilt_2_4/azure_layout_high_res_formulas/records.jsonl': Value('int64'), 'parses/tilt_2_4/cohere_parse_v5/records.jsonl': Value('int64'), 'parses/tilt_2_4/datalab_accurate/records.jsonl': Value('int64'), 'parses/tilt_2_4/extend_performance_agentic_formulas/records.jsonl': Value('int64'), 'parses/tilt_2_4/llamaparse_agentic_plus/records.jsonl': Value('int64'), 'parses/tilt_2_4/mistral_ocr_4_1/records.jsonl': Value('int64'), 'parses/tilt_2_4/pulse_ultra_2/records.jsonl': Value('int64'), 'parses/tilt_2_4/reducto_r1/records.jsonl': Value('int64')}}}, 'native_backfill': Value('string'), 'native_parse_matches': Value('string'), 'native_validation': Value('string')}
because column names don't match
Traceback: 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 361, 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 2381, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2309, in cast_table_to_schema
raise CastError(
...<3 lines>...
)
datasets.table.CastError: Couldn't cast
compared: int64
match: int64
file_bytes_match: int64
markdown_match: int64
included_native_rows: int64
excluded_native_rows: list<item: null>
child 0, item: null
parse_mismatches: list<item: null>
child 0, item: null
aggregate_files: int64
hash_definition: string
snapshot: string
benchmarks: struct<dp_bench: struct<files: int64, bytes: int64, records: struct<judge/clean/gpt-6.1-sol-medium-r (... 26980 chars omitted)
child 0, dp_bench: struct<files: int64, bytes: int64, records: struct<judge/clean/gpt-6.1-sol-medium-r0/records.jsonl: (... 2128 chars omitted)
child 0, files: int64
child 1, bytes: int64
child 2, records: struct<judge/clean/gpt-6.1-sol-medium-r0/records.jsonl: int64, judge/clean/gpt-6.1-sol-medium-r0/req (... 2083 chars omitted)
child 0, judge/clean/gpt-6.1-sol-medium-r0/records.jsonl: int64
child 1, judge/clean/gpt-6.1-sol-medium-r0/requests.jsonl: int64
child 2, judge/clean/gpt-6.1-sol-medium-r1/records.jsonl: int64
child 3, judge/clean/gpt-6.1-sol-medium-r1/requests.jsonl: int64
child 4, judge/tilt_2_4/gpt-6.1-sol-medium-r0/records.jsonl: int64
child 5, judge/tilt_2_4/gpt-6.1-sol-medium-r0/requests.jsonl: int64
child 6, native/clean/azure_layout_high_res_formulas/scores.jsonl: int64
child 7, native/clean/cohere_parse_v5/scores.jsonl: int64
child 8, native/clean/datalab_accurate/scores.jsonl: int64
child 9, native/clean/extend_performance
...
ocr_4_1/scores.jsonl: int64
child 20, native/tilt_2_4/pulse_ultra_2/scores.jsonl: int64
child 21, native/tilt_2_4/reducto_r1/scores.jsonl: int64
child 22, pages.jsonl: int64
child 23, parses/clean/azure_layout_high_res_formulas/records.jsonl: int64
child 24, parses/clean/cohere_parse_v5/records.jsonl: int64
child 25, parses/clean/datalab_accurate/records.jsonl: int64
child 26, parses/clean/extend_performance_agentic_formulas/records.jsonl: int64
child 27, parses/clean/llamaparse_agentic_plus/records.jsonl: int64
child 28, parses/clean/mistral_ocr_4_1/records.jsonl: int64
child 29, parses/clean/pulse_ultra_2/records.jsonl: int64
child 30, parses/clean/reducto_r1/records.jsonl: int64
child 31, parses/tilt_2_4/azure_layout_high_res_formulas/records.jsonl: int64
child 32, parses/tilt_2_4/cohere_parse_v5/records.jsonl: int64
child 33, parses/tilt_2_4/datalab_accurate/records.jsonl: int64
child 34, parses/tilt_2_4/extend_performance_agentic_formulas/records.jsonl: int64
child 35, parses/tilt_2_4/llamaparse_agentic_plus/records.jsonl: int64
child 36, parses/tilt_2_4/mistral_ocr_4_1/records.jsonl: int64
child 37, parses/tilt_2_4/pulse_ultra_2/records.jsonl: int64
child 38, parses/tilt_2_4/reducto_r1/records.jsonl: int64
native_validation: string
native_parse_matches: string
native_backfill: string
to
{'snapshot': Value('string'), 'benchmarks': {'dp_bench': {'files': Value('int64'), 'bytes': Value('int64'), 'records': {'judge/clean/gpt-6.1-sol-medium-r0/records.jsonl': Value('int64'), 'judge/clean/gpt-6.1-sol-medium-r0/requests.jsonl': Value('int64'), 'judge/clean/gpt-6.1-sol-medium-r1/records.jsonl': Value('int64'), 'judge/clean/gpt-6.1-sol-medium-r1/requests.jsonl': Value('int64'), 'judge/tilt_2_4/gpt-6.1-sol-medium-r0/records.jsonl': Value('int64'), 'judge/tilt_2_4/gpt-6.1-sol-medium-r0/requests.jsonl': Value('int64'), 'native/clean/azure_layout_high_res_formulas/scores.jsonl': Value('int64'), 'native/clean/cohere_parse_v5/scores.jsonl': Value('int64'), 'native/clean/datalab_accurate/scores.jsonl': Value('int64'), 'native/clean/extend_performance_agentic_formulas/scores.jsonl': Value('int64'), 'native/clean/llamaparse_agentic_plus/scores.jsonl': Value('int64'), 'native/clean/mistral_ocr_4_1/scores.jsonl': Value('int64'), 'native/clean/pulse_ultra_2/scores.jsonl': Value('int64'), 'native/clean/reducto_r1/scores.jsonl': Value('int64'), 'native/tilt_2_4/azure_layout_high_res_formulas/scores.jsonl': Value('int64'), 'native/tilt_2_4/cohere_parse_v5/scores.jsonl': Value('int64'), 'native/tilt_2_4/datalab_accurate/scores.jsonl': Value('int64'), 'native/tilt_2_4/extend_performance_agentic_formulas/scores.jsonl': Value('int64'), 'native/tilt_2_4/llamaparse_agentic_plus/scores.jsonl': Value('int64'), 'native/tilt_2_4/mistral_ocr_4_1/scores.jsonl': Value('int64'), 'native/tilt_2_4
...
4'), 'native/tilt_2_4/llamaparse_agentic_plus/scores.jsonl': Value('int64'), 'native/tilt_2_4/mistral_ocr_4_1/scores.jsonl': Value('int64'), 'native/tilt_2_4/pulse_ultra_2/scores.jsonl': Value('int64'), 'native/tilt_2_4/reducto_r1/scores.jsonl': Value('int64'), 'pages.jsonl': Value('int64'), 'parses/clean/azure_layout_high_res_formulas/records.jsonl': Value('int64'), 'parses/clean/cohere_parse_v5/records.jsonl': Value('int64'), 'parses/clean/datalab_accurate/records.jsonl': Value('int64'), 'parses/clean/extend_performance_agentic_formulas/records.jsonl': Value('int64'), 'parses/clean/llamaparse_agentic_plus/records.jsonl': Value('int64'), 'parses/clean/mistral_ocr_4_1/records.jsonl': Value('int64'), 'parses/clean/pulse_ultra_2/records.jsonl': Value('int64'), 'parses/clean/reducto_r1/records.jsonl': Value('int64'), 'parses/tilt_2_4/azure_layout_high_res_formulas/records.jsonl': Value('int64'), 'parses/tilt_2_4/cohere_parse_v5/records.jsonl': Value('int64'), 'parses/tilt_2_4/datalab_accurate/records.jsonl': Value('int64'), 'parses/tilt_2_4/extend_performance_agentic_formulas/records.jsonl': Value('int64'), 'parses/tilt_2_4/llamaparse_agentic_plus/records.jsonl': Value('int64'), 'parses/tilt_2_4/mistral_ocr_4_1/records.jsonl': Value('int64'), 'parses/tilt_2_4/pulse_ultra_2/records.jsonl': Value('int64'), 'parses/tilt_2_4/reducto_r1/records.jsonl': Value('int64')}}}, 'native_backfill': Value('string'), 'native_parse_matches': Value('string'), 'native_validation': Value('string')}
because column names don't matchNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
BenchBench runs
Saved outputs for BenchBench's 12-benchmark study of eight commercial parsers. The
primary source is the Chandra 2.7 final refresh, identified by 20261116T220719Z.
The code repository is reductoai/benchbench.
This dataset is private pending release review.
Scope
Each benchmark has the saved sample manifest, parser outputs with sanitized vendor
responses, and primary judge verdicts and configuration. Conditions are clean and
tilt_2_4. The tilt protocol adds a 2% white margin and rotates 2-4 degrees in either
direction. The seed is 20261107. Page images and full ground-truth workspaces are
not included. Recorded diagnostics can quote benchmark-derived text.
The small summary/ snapshot contains the code repository's published
analysis inputs, including its existing sample metadata and historical summary tables.
The primary judge is PAIR, based on the pinned Parse Judge compare-v4 image prompt, run with GPT-6.1 Sol at medium effort. It scores all eight outputs together with vendor names hidden. Clean averages r0 and r1; tilt uses r0. The judge sees the clean reference image. Verdict records retain the model, prompt hash, preset order, blinded labels, usage, and parse hashes where the source recorded them. Sanitized HTTP judge request and response exchanges are archived separately. Ordinary image links are retained; embedded binary payloads become artifact references. Page images must be obtained upstream.
Datalab uses accurate mode with image extraction and captions enabled. This refresh
replaced Datalab's parses and rejudged all eight vendors together. The other seven
vendors' parses were reused. Datalab native records come from the final refresh. For the
seven reused presets, clean/tilt native records are backfilled from historical benchmark
runs only where the historical and final parse content hashes match. The hash includes
parser status, Markdown, raw response, harness-error status and image hash.
NATIVE_PARSE_MATCHES.jsonl records every comparison; NATIVE_BACKFILL.json gives counts
and any exclusions. The archive now supports checking all 192 primary native means.
It still cannot recompute every historical appendix interval from per-page data.
Historical scan, perspective, Gemini, minimal and terse raw runs are not included.
Their summary values predate the Datalab refresh.
Layout
Each benchmark uses its exact ID as a top-level folder:
dp_benchidp_leaderboardmdpbenchocrflux_bencholmocr_benchomnidocbenchomniparsebenchparsebench2puredocbenchrd_tablebenchreal5_omnidocbenchscore_bench
Paths within a benchmark:
| Path | Contents |
|---|---|
pages.jsonl, run.json |
Sample IDs, seed, source revisions and run settings |
parses/<condition>/<preset>/records.jsonl |
One sanitized parse record per line, including vendor response fields |
judge/<condition>/<judge-tag>/records.jsonl |
One saved judge record per line, including blinded mapping and verdicts |
judge/<condition>/<judge-tag>/requests.jsonl |
Sanitized HTTP request and response exchanges, keyed by page |
native/<condition>/<preset>/scores.jsonl |
Native scores, subscores and recorded diagnostics per eligible page |
native/**/official.json |
Benchmark-level native aggregation where present |
native_work/**/official.json |
Other adapter headline aggregations where present |
native_work/**/lines.jsonl |
OmniParseBench test types and results only, without test contents |
summary/ mirrors the code repository's results/ directory after the same sanitization.
INVENTORY.json records exact per-benchmark sizes and record counts. SHA256SUMS covers
the published files. restore_runs.py expands shards to the harness's file-per-page layout.
The JSONL shards reduce small-file overhead. They are not directly readable by the
harness until expanded.
Use with the code repository
Download this dataset, then run the helper from the download directory:
python restore_runs.py ./expanded
export BENCHBENCH_RUNS="$PWD/expanded"
In an installed code checkout, generate a report without API calls:
B=score_bench
P="reducto_r1 llamaparse_agentic_plus datalab_accurate extend_performance_agentic_formulas mistral_ocr_4_1 azure_layout_high_res_formulas pulse_ultra_2 cohere_parse_v5"
benchbench report "$BENCHBENCH_RUNS/$B" --benchmark "$B" --presets $P --conditions clean tilt_2_4 --judge gpt-6.1-sol --repeats 2
PAIR scores can be checked from the full verdicts. The archive includes clean/tilt native records for all eight presets. The generic report's page-score summaries are not every benchmark's native headline aggregation. Parser and judge reruns would make paid calls and need fresh images and credentials. Expansion does not fetch either.
To regenerate the saved summary tables from this package, use the code checkout:
python analysis/make_results_md.py --results /path/to/download/summary --output ./RESULTS.md
BENCHBENCH_RESULTS=/path/to/download/summary python analysis/headlines.py
The published summary has 96 parser-benchmark pairs. Native winners change after tilt on 6/12 benchmarks, PAIR winners on 2/12. Median relative losses are 2.2% and 4.0%. The percentage lead-versus-loss comparison is 9/12. Its raw-score-point variant is 8/12. See the code repository's generated results and methodology for definitions and caveats.
Run python validate_native.py from the download directory to recompute all 192 primary
native means. NATIVE_VALIDATION.json records each comparison and its aggregation.
Legacy inputs were saved to two or three decimals in summary/native_rotated_corrected.csv.
Refreshed Datalab values retain more precision. Comparisons respect those saved resolutions.
OmniDocBench uses the retained official pooled text/table/formula components, not a
mean of page scores. This checks aggregation of recorded scorer results, not a new
native scoring run against benchmark ground truth.
ParseBench-2 retains a historical aggregation difference. The seven reused presets use
the historical 480-page complete-case cohort and arithmetic page means, as in the code
repository's analysis/build_data.py. parsebench2/native/cohorts.json preserves the
membership and source-shard hashes. Only membership from the old Datalab scores is used;
those old scores are not included. Refreshed Datalab uses subset-weighted dimension means
without layout. The archive preserves these conventions rather than harmonizing them.
PureDocBench also has a historical aggregation difference. The seven reused presets use
pooled formula/table components in native/<condition>/<preset>/pooled.json. Refreshed
Datalab uses page-averaged task components.
Unresolved validation differences
The checker evaluates 192 native means. Ten do not match the saved values at source
precision. NATIVE_VALIDATION.json lists them. Do not treat this as a fully reconciled
raw-data release. The source also lacks native rows for 41 matched MDPBench parse pages;
NATIVE_MISSING_ROWS.jsonl lists them. No native rows were excluded for parse mismatches.
For refreshed ParseBench-2, the headline can be reconstructed from the recorded official
dimension means. Its recorded Tables dimension does not equal the weighted mean of its
recorded subset scores in either condition. NATIVE_COMPONENT_CHECKS.json records these
two inconsistencies. The archive preserves the source values; it does not repair them.
Sanitization
Ordinary text and metadata are preserved, including names, words such as internal,
bucket names, hostnames, local paths and links. No text-redaction placeholders are
inserted. Literal redaction markers already present in parser or judge outputs remain
unchanged. Credential-bearing fields such as Authorization and x-api-key are dropped.
URL fields retain their query encoding except that X-Amz-Signature, X-Amz-Credential
and X-Amz-Security-Token query parameters are removed. Other parameters remain.
Credential-shaped strings inside document, parser or judge text are not edited; they
require separate owner review before release. Page images and embedded image/PDF
binary payloads remain omitted. Textual requests, responses, reasoning and native
diagnostics are otherwise preserved. Recorded scores are unchanged.
SANITIZATION.json records the credential and binary removals.
Native workspaces, scorer inputs, ground-truth rules, operational scripts, logs, page images and extracted image payloads are omitted. Native score records retain original test identifiers and explanations. Selected OmniParseBench native-work files retain test family and result, not test contents.
Licenses and limits
Reducto's judge verdicts and new dataset files are licensed CC BY 4.0. Vendor outputs and benchmark-derived content keep their own provider and source terms. The YAML license describes our contributions, not a grant over third-party material. The code repository lists benchmark licenses. These outputs are provided for research. No right to redistribute source page images is granted.
Reducto sells a parser in the study, wrote the judge prompt, and built RD-TableBench. LlamaIndex built ParseBench-2. Datalab built OmniParseBench. Both sell parsers in the study. The judge's figure-description rule and the historical figure-free subset check are disclosed in the code repository. API drift, finite samples, historical coverage gaps and binary omissions limit what this archive can establish. Do not treat its ranks as a universal parser leaderboard.
Citation
Reducto. BenchBench: testing document-parsing benchmarks. 2026.
See CITATION.cff in the code repository.
Sizes
Exact byte and record counts are in INVENTORY.json. MiB uses 1,048,576 bytes.
| Benchmark ID | Artifact files | Sanitized MiB |
|---|---|---|
dp_bench |
56 | 110.9 |
score_bench |
40 | 175.1 |
olmocr_bench |
40 | 426.3 |
omniparsebench |
42 | 459.8 |
ocrflux_bench |
40 | 362.0 |
idp_leaderboard |
40 | 149.6 |
mdpbench |
40 | 494.3 |
rd_tablebench |
40 | 234.6 |
puredocbench |
54 | 790.2 |
omnidocbench |
56 | 450.3 |
real5_omnidocbench |
56 | 477.1 |
parsebench2 |
57 | 558.8 |
Sizes above cover benchmark artifacts, not the small summary snapshot and metadata.
- Downloads last month
- 51