Datasets:

sleepyheeler
/

okra-parsestream-bench

Error code:   StreamingRowsError
Exception:    CastError
Message:      Couldn't cast
id: string
title: string
source: string
license: string
pdf_url: string
pdf_path: string
doc_class: list<item: string>
  child 0, item: string
pages: int64
pdf_bytes: int64
gt: struct<has_text_layer: bool, scanned: bool, needs_vision: bool, figures: list<item: struct<page: int (... 239 chars omitted)
  child 0, has_text_layer: bool
  child 1, scanned: bool
  child 2, needs_vision: bool
  child 3, figures: list<item: struct<page: int64, bbox: list<item: double>, kind: string, raster_hits: int64, raster_co (... 48 chars omitted)
      child 0, item: struct<page: int64, bbox: list<item: double>, kind: string, raster_hits: int64, raster_cover: double (... 36 chars omitted)
          child 0, page: int64
          child 1, bbox: list<item: double>
              child 0, item: double
          child 2, kind: string
          child 3, raster_hits: int64
          child 4, raster_cover: double
          child 5, vector_ops: int64
          child 6, fig_id: string
  child 4, gt_source: string
  child 5, n_figures: int64
  child 6, n_raster: int64
  child 7, n_vector: int64
  child 8, n_in_page_raster: int64
  child 9, n_unresolved: int64
gt_note: string
gt_evidence: struct<median_chars_per_page: int64, chars_per_page_sampled: int64, page_image_frac: double, garbage (... 451 chars omitted)
  child 0, median_chars_per_page: int64
  child 1, chars_per_page_sampled: int64
  child 2, page_image_frac: double
  child 3, garbage_char_ratio: double
  child 4, garbage_flag: bool
  child 5, inspector_type: string
  child 6, inspector_confidence: double
  child 7, inspector_pages_needing_ocr: int64
  child 8, inspector_says_scanned: bool
  child 9, liteparse: null
  child 10, n_image_placements: int64
  child 11, n_figure_candidates: int64
  child 12, n_inspector_image_placeholders: int64
  child 13, n_image_xobjects: int64
  child 14, n_cmyk_jpeg: int64
  child 15, n_smask: int64
  child 16, median_vector_ops_per_page: int64
  child 17, run_status: string
  child 18, t_classify_ms: double
  child 19, t_scan_ms: double
  child 20, t_inspect_ms: double
disagreements: list<item: string>
  child 0, item: string
page_length: int64
full_confidence: string
stress_tags: list<item: string>
  child 0, item: string
gold_html_path: string
gold_html_url: string
doc_type: string
xml_url: string
to
{'id': Value('string'), 'source': Value('string'), 'title': Value('string'), 'doc_type': Value('string'), 'pdf_path': Value('string'), 'gold_html_path': Value('string'), 'pdf_url': Value('string'), 'gold_html_url': Value('string'), 'full_confidence': Value('string'), 'license': Value('string'), 'stress_tags': List(Value('string')), 'xml_url': Value('string'), 'page_length': Value('int64'), 'pdf_bytes': Value('int64')}
because column names don't match
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 2306, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              id: string
              title: string
              source: string
              license: string
              pdf_url: string
              pdf_path: string
              doc_class: list<item: string>
                child 0, item: string
              pages: int64
              pdf_bytes: int64
              gt: struct<has_text_layer: bool, scanned: bool, needs_vision: bool, figures: list<item: struct<page: int (... 239 chars omitted)
                child 0, has_text_layer: bool
                child 1, scanned: bool
                child 2, needs_vision: bool
                child 3, figures: list<item: struct<page: int64, bbox: list<item: double>, kind: string, raster_hits: int64, raster_co (... 48 chars omitted)
                    child 0, item: struct<page: int64, bbox: list<item: double>, kind: string, raster_hits: int64, raster_cover: double (... 36 chars omitted)
                        child 0, page: int64
                        child 1, bbox: list<item: double>
                            child 0, item: double
                        child 2, kind: string
                        child 3, raster_hits: int64
                        child 4, raster_cover: double
                        child 5, vector_ops: int64
                        child 6, fig_id: string
                child 4, gt_source: string
                child 5, n_figures: int64
                child 6, n_raster: int64
                child 7, n_vector: int64
                child 8, n_in_page_raster: int64
                child 9, n_unresolved: int64
              gt_note: string
              gt_evidence: struct<median_chars_per_page: int64, chars_per_page_sampled: int64, page_image_frac: double, garbage (... 451 chars omitted)
                child 0, median_chars_per_page: int64
                child 1, chars_per_page_sampled: int64
                child 2, page_image_frac: double
                child 3, garbage_char_ratio: double
                child 4, garbage_flag: bool
                child 5, inspector_type: string
                child 6, inspector_confidence: double
                child 7, inspector_pages_needing_ocr: int64
                child 8, inspector_says_scanned: bool
                child 9, liteparse: null
                child 10, n_image_placements: int64
                child 11, n_figure_candidates: int64
                child 12, n_inspector_image_placeholders: int64
                child 13, n_image_xobjects: int64
                child 14, n_cmyk_jpeg: int64
                child 15, n_smask: int64
                child 16, median_vector_ops_per_page: int64
                child 17, run_status: string
                child 18, t_classify_ms: double
                child 19, t_scan_ms: double
                child 20, t_inspect_ms: double
              disagreements: list<item: string>
                child 0, item: string
              page_length: int64
              full_confidence: string
              stress_tags: list<item: string>
                child 0, item: string
              gold_html_path: string
              gold_html_url: string
              doc_type: string
              xml_url: string
              to
              {'id': Value('string'), 'source': Value('string'), 'title': Value('string'), 'doc_type': Value('string'), 'pdf_path': Value('string'), 'gold_html_path': Value('string'), 'pdf_url': Value('string'), 'gold_html_url': Value('string'), 'full_confidence': Value('string'), 'license': Value('string'), 'stress_tags': List(Value('string')), 'xml_url': Value('string'), 'page_length': Value('int64'), 'pdf_bytes': Value('int64')}
              because column names don't match

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.

ParseStream — does a generative PDF→HTML stream behave like a real website?

Benchmark harness + recorded runs for okraPDF's live twin (okrapdf.com/s/:id): a page whose "load" is a generative VLM stream, not a static asset. Existing PDF→HTML benchmarks score only final-artifact fidelity; web-perf tools (Lighthouse/CrUX) assume the bytes already exist. ParseStream scores the load trajectory and the HTTP surface of a document that is being written while you read it.

Metrics

Lane	Instrument	Scores
Lane 0	`harness/lane0_api_surface.mjs`	~22 weighted API-surface checks: JSON create receipt, no-JS text-rendition coverage, ETag/304 revalidation, SSE resume via `?after=`/`Last-Event-ID` mid-stream, wire quiet-gap (heartbeats), 425+Retry-After download handshake, HEAD support, compression
Lane 1	`harness/lane1_tap.mjs`	client-receive SSE timing + full `okra.live_html.v1` contract validation: TTFB, TTFE, TTFC, cadence p50/p95, max-stall, smoothness, TTLB, figure-resolve
Trajectory	`harness/trajectory.mjs`	trajectory regret (normalized area between the byte-fill curve and an instant-complete page: 0 ≈ static page, →1 ≈ spinner-then-dump), fill-50/90/100, figure-tail share, stall attribution (prep / cadence / figure-lane)
Aggregate	`harness/where_to_improve.mjs`	gates → ranked findings, each naming its fix seam

Headline results (2026-08-11, engine E1 single-call, worker `0fa8d159`)

API surface: 22/22 after the CX.1.8f fix round (runs/pre-deploy/ holds the before-state: no-JS coverage 0.006, no validators, no HEAD, 303-only create).
Trajectory regret p50 0.597 (gate ≤ 0.35) — the stream is still too back-loaded.
Figure-lane overhang p95 = 71% of load — text done at ~19s, figures at ~68s on the worst doc.
TTFC p50 3.4s (gate ≤ 2.5s) — floored by whole-doc VLM first-token; the per-page-parallel engine (E2) is the named fix.

Reproduce

cd harness
# score the recorded runs offline (no network):
node trajectory.mjs ../runs/*.events.jsonl
node where_to_improve.mjs --runs ../runs
# probe the live surface (creates a real session; be polite — the endpoint is rate-limited):
node lane0_api_surface.mjs --demo --base https://okra-a11y-agent.steventsao.workers.dev
node lane1_tap.mjs --url <public-pdf-url> --id mydoc --base https://okrapdf.com

runs/*.events.jsonl are timestamped SSE event logs — state(t) = apply(events ≤ t) replays the DOM trajectory offline. *.metrics.json / *.trajectory.json / *.lane0.json are the scored outputs.

Corpus & rights

Source documents are government publications with authoritative accessible-HTML twins (manifest/govtwin.jsonl): GOV.UK publications under the Open Government Licence v3.0 (contains public sector information licensed under OGL v3.0) and US Federal Register documents (US federal works, 17 USC §105). runs/*.final.html are machine-generated renditions of those documents produced by the system under test; they inherit the source licences. Harness code (harness/*.mjs) is provided under Apache-2.0.

Spec with the full metric design + research framing: reports/streaming-vitality-bench.md
Current ranked findings: reports/WHERE-TO-IMPROVE.md
Sibling dataset: sleepyheeler/okra-gdp-pdf-bench-traces

Image-layer lane (added 2026-08-12)

harness/lane_imglayer.mjs + runs-imglayer/: can direct image-layer extraction (Firecrawl pdf-inspector placeholders + MuPDF byte extraction) replace the VLM twin's rasterize-after-parse figure lane? Verdict: routed fast path, not a replacement — raster-native worst case 36.8s -> 305ms (39/39 figures), vector-only docs keep the rasterizer, routing verdict ~14ms. Traps (CMYK SOF gating, /SMask compositing, wasm-view detachment, bottom-left bbox origin) documented in runs-imglayer/SUMMARY.md.

Graded standings — corpus v2 (added 2026-08-12)

manifest/gradeset.jsonl (19 docs, 8 stratified classes incl. 4 scanned) + harness/{build_gradeset,route,grade}.mjs -> reports/{STANDING,GRADES}.md. Grades the ROUTING POLICY as a first-class dimension: doc-level 19/19 (weak evidence — all public-record scans carry publisher OCR; true no-text scan = missing cell), figure routing micro-F1 1.000 with a measured >=20% area-coverage gate, figure delivery 562x-2549x, vitality 0.283 (F). Ground-truth adjudication renders in runs-grades/adjudication/.

Downloads last month: 198

Total file size:

12.1 MB