parsebench2-viewer / README.md
Hashir621's picture
Add VLM prompts (index/prompts.json)
f3edd56 verified
|
Raw History Blame Contribute Delete
15.1 kB
metadata
license: apache-2.0
pretty_name: ParseBench 2 Viewer (unofficial)
tags:
  - document-parsing
  - pdf
  - benchmark
  - parsebench
configs:
  - config_name: pipelines
    data_files: pipelines.parquet
  - config_name: scores
    data_files: metrics/scores.parquet
  - config_name: page_dimension_scores
    data_files: metrics/page_dimension_scores.parquet
  - config_name: events
    data_files: metrics/events/*.parquet
  - config_name: layout_elements
    data_files: metrics/layout_elements.parquet
  - config_name: layout_predictions
    data_files: metrics/layout_predictions.parquet
  - config_name: pages
    data_files: pages.parquet
  - config_name: groundtruth
    data_files: groundtruth/groundtruth.parquet
  - config_name: images
    data_files: images/images.parquet
  - config_name: slots
    data_files: slots.parquet

ParseBench 2 Viewer (unofficial)

An unofficial community dataset derived from ParseBench by LlamaIndex, licensed Apache 2.0 like its sources. It collects the per-page results of the 56 parsers on the ParseBench 2 leaderboard (4,317 test pages) in one place: every score and scoring event, per-page latency and cost, the ground truth, every parser's output, and a thumbnail and preview of every page. It is laid out so a static website can read it directly: side-by-side comparisons, each parser's lowest-scoring pages per dimension, and parser-vs-parser head-to-heads.

No parser or scorer was re-run and no score was changed; the data is only reshaped and joined.

Sources (pinned)

Source Repo Commit
Parser outputs, latency, cost llamaindex/ParseBench-2-results 21ecc920c6969f69353d4be51d27a4346979f645
Ground truth and page documents llamaindex/ParseBench-2 2d9a2344660647e87dc272458ea1978428f895eb
Scorer and leaderboard-v2.csv run-llama/ParseBench d7efdbc864e2e55f9793063a90936123fc769004

Every input was checked against its SHA-256 before use; see viewer_manifest.json and validation_report.json.

Scores vs. the leaderboard

Recomputed dimension and overall scores equal each parser's scorer output to within 1e-9. They match the published leaderboard on Text, Formatting, Tables, Charts and Layout. Forms is 0.01–0.91 points higher than the leaderboard because it uses the newer ground truth and scorer (delta_forms in pipelines.parquet).

A dimension score is the weighted mean of its slots (slots.parquet); a slot is one metric averaged over one subset's pages. Pages where the slot's metric is null (no formatting in the reference, for example) are left out of that average; pages with no published output keep their metrics, scored as 0. Overall is the plain mean of the six dimensions.

Files

Path Rows / files What it is
pipelines.parquet 56 One row per parser: leaderboard values, recomputed scores (0–100), deltas, cost, bbox_frame
slots.parquet 14 Which metric counts for which subset and dimension, with weights
pages.parquet 4,317 Page index: subset, dimensions, pinned document URL and hash, page size, image paths
metrics/scores.parquet 241,752 One row per (parser, page): all ~45 metrics, severity counts, layout totals, latency, cost, tokens
metrics/page_dimension_scores.parquet 268,632 One row per (parser, page, dimension): the counted score, points_lost, rank among parsers, page mean/min/max
metrics/events/<metric>.parquet 3.06M Every scoring event for text_next, formatting_v2, table_next, chart_next, forms_v2
metrics/layout_elements.parquet 930,440 Layout result for every reference element
metrics/layout_predictions.parquet 95,000 Unmatched layout predictions and their penalties
groundtruth/groundtruth.parquet 4,317 Full reference per page (metadata_json, test_rules_json as compact JSON)
documents/<subset>/<page_key>.pdf 4,317 The page document itself (4,296 PDFs and 21 .jpg), copied byte-for-byte from the pinned ground-truth repo
images/images.parquet 4,317 Image sizes, paths and atlas positions
images/thumbs/<subset>/<page_key>.webp 4,317 240 px wide
images/previews/<subset>/<page_key>.webp 4,317 1,200 px wide
images/atlas/<subset>.webp + .json 13 + 13 Every thumbnail of a subset on one sheet
bundles/metrics/<subset>/<page_key>.json 4,317 All 56 parsers' full metric lists for one page
bundles/outputs/<subset>/<page_key>.json 4,317 All 56 parsers' markdown for one page, with its pinned source URL
bundles/layout_boxes/layout/<page_key>.json 535 All 56 parsers' layout boxes for one layout page
bundles/groundtruth/<subset>/<page_key>.json 4,317 Full ground-truth record, metadata parsed
index/*.json 66 Precomputed JSON for static sites (below)
index/prompts.json 26 pipelines The exact prompt each VLM pipeline sends, with links to its source lines (below)

page_key is the first 16 hex characters of sha256(test_id); file paths use it because test ids contain slashes, spaces and parentheses. Each page's document is in this repo at pages.document_file (SHA-256 in pages.document_sha256); pages.document_url also points at the same file at the pinned commit of llamaindex/ParseBench-2.

Using it from a static site

Every file is served with CORS and HTTP range support at https://huggingface.co/datasets/Hashir621/parsebench2-viewer/resolve/<commit>/<path>. Pin <commit> so the data never changes under you. Parquet files can be queried in the browser with DuckDB-WASM or hyparquet; the JSON files need nothing.

  • index/site.json: the entry point. Parsers (rank, scores, slot scores, latency/cost summary), dimensions, slots, subsets with atlas geometry, and path templates for every per-page file.
  • index/pages.json: every page in columnar form (columns + rows) with thumbnail size and atlas position.
  • index/dimensions/<dimension>.json: a page × parser matrix of the counted score, points_lost and critical-event count, plus per-page mean/min/max/std. Any A-vs-B comparison or "where do parsers disagree most" view is arithmetic on this file.
  • index/parsers/<pipeline>.json: for each dimension, every scored page sorted lowest score first, with points_lost, the parser's rank on that page and the page mean.
  • index/pairwise.json: per dimension, 56 × 56 head-to-head wins, ties and mean score difference.

points_lost is 100 × slot_weight / pages_in_slot × (1 − score): how many points that page cost the parser's dimension score. A parser's points_lost over a dimension sums to exactly 100 − dimension score, so it ranks pages by real impact even when a dimension mixes subsets of different weight (Formatting, Tables, Forms). The text subset belongs to two dimensions (Text and Formatting), so its pages appear under both.

A page view needs documents/… (the PDF, e.g. for PDF.js; or images/previews/… as a lighter stand-in), bundles/outputs/… (markdown), bundles/metrics/… (scores and events: why a page scored what it did) and bundles/groundtruth/… (reference); add bundles/layout_boxes/… for box overlays on layout pages. For a two-parser comparison with a tie threshold, compare the two columns of index/dimensions/<dimension>.json (or page_dimension_scores.parquet); index/pairwise.json counts exact ties only.

VLM prompts

index/prompts.json holds the prompt each of 26 pipelines sends, as written in run-llama/ParseBench at the pinned commit: 20 general vision-language models (Claude, GPT, Gemini, Muse, Qwen, Gemma), 5 OCR models that take a single fixed prompt (dots.ocr, Chandra, LightOnOCR, Jina), and the custom instructions Pulse Ultra 2 passes to its API. Each entry lists its messages in order (system, user, or the API field name), a note on how the page is sent, and a source link to the exact lines that define the text.

The prompts were extracted statically: the ParseBench source was parsed, and only string constants, joins, .replace() and f-strings over constants were evaluated; no ParseBench code was run. The Claude Opus and Haiku 5.5 image-mode prompts end with one sentence stating the page image's size; it appears as a template with {width} and {height}. Pipelines whose prompts are multi-stage or built into the model's SDK (HunyuanOCR, PaddleOCR-VL, KDL, MinerU, Infinity, MiniCPM, TeleOCR, Surya) and document-parsing APIs without custom instructions are not listed. site.json marks each pipeline with has_prompt.

Event severities

formatting_v2, table_next and forms_v2 events carry their own severity (severity_source = stored). text_next and chart_next events do not, so it is set from the scorer's rules (severity_source = mapped):

  • text_next — critical: checkbox flipped, invented passage, meaning flipped, name changed, negator in added text, negator in changed text, numbers swapped, omitted passage, sign changed, wrong number or identifier; major: blocks out of order, identifier garbled, inserted words, omitted, repeated text, sentences cut, wrong words; minor: accent only, look-alike characters, number written differently, spelling, words split or joined; info (explains a loss, not charged on its own): critical-token sequence differs, reading order.
  • chart_next — critical: chart missing, series missing, wrong value, wrong value (scale); major: category label reworded, category label truncated, category missing, derived total, invented series or category, missing, point missing, repeated point, repeated row, series renamed, title missing, value off, wrong title; minor: category axis mislabeled, colour groups merged, label typo, range, range instead of value, series name reworded.
  • legend lost is critical when the chart has 2 or more series and major otherwise; the event does not say which, so its severity is null.
  • title missing and series name reworded are not named in the original build notes; their severities come from chart_next.py at the pinned commit.

For blocks out of order, sentences cut, look-alike characters and the count form of number written differently, reference holds a count, not text. Invented numbers in inserted text add to the critical count without an event of their own, so text_next__n_critical can exceed the critical events listed. Table and forms events are normalized: type → kind, expected → reference.

Coordinate frames

The two kinds of boxes use different frames and are kept as published:

  • Ground-truth layout rules (test_rules entries with a bbox): [x, y, w, h] as fractions of the page, 0–1.
  • Parser layout boxes (layout_pages[].items[].bbox): {x, y, w, h, label, confidence, r}; each page also has width and height. pipelines.bbox_frame says which frame a parser uses: fraction for 50 parsers, page_units (divide by the page's width/height) for llamaparse_agentic_plus, llamaparse_agentic, llamaparse_cost_effective, paddleocr_vl_1_6_pipeline, falconocr_pipeline, and null for jinaocr_vllm, which publishes no boxes. Fractional boxes can overshoot the page edge slightly (up to ~1.5% of coordinates are above 1 for some parsers).

Pages with no output

44 (parser, page) rows have no published output: no record in the results files, so markdown_sha256, the markdown and the timing columns are null. Their metrics are kept, scored as 0.

Parser Rows Pages
qwen3_8_27b_thinking_parse_with_layout 16 charts/ac8b3538-en_p36, charts/ac8b3538-en_p53, dense_tables/dense_table_155, dense_tables/dense_table_156, dense_tables/dense_table_158, dense_tables/dense_table_159, dense_tables/dense_table_167, dense_tables/dense_table_168, dense_tables/dense_table_253, forms/filled__france--47_cerfa_16041--p2__typed, forms/filled__us--27_Schedule-E--p1__typed, insurance_filing_tables/08451152a59275c1_p1, insurance_filing_tables/1a458e0324657e28_p16, insurance_filing_tables/4712995c09c36224_p28, tables/p29_page1, tables/synthetic_invoice_page2
openai_gpt_5_6_luna_reasoning_low_parse_with_layout 5 dense_tables/dense_table_153, forms/blank__japan--01_kakutei_shinkokusho--p2__blank, forms/filled__japan--36_junkakutei_fuhyo--p1__typed, insurance_filing_tables/a6051a8405563b92_p9, text/text_ocr__ref
openai_gpt_5_6_terra_reasoning_low_parse_with_layout 5 dense_tables/dense_table_153, forms/blank__japan--01_kakutei_shinkokusho--p2__blank, forms/filled__japan--36_junkakutei_fuhyo--p1__typed, insurance_filing_tables/a6051a8405563b92_p9, text/text_ocr__ref
openai_gpt_6_luna_reasoning_low_parse_with_layout 5 dense_tables/dense_table_153, forms/blank__japan--01_kakutei_shinkokusho--p2__blank, forms/filled__japan--36_junkakutei_fuhyo--p1__typed, insurance_filing_tables/a6051a8405563b92_p9, text/text_ocr__ref
pulse_ultra_2 3 charts/sri-sigma-natural-catastrophes-1-2025_p4, forms/filled__us--02_1040--p2__typed, text/text_sparse__notes
databricks_ai_parse 2 dense_tables/dense_table_214, insurance_filing_tables/a6051a8405563b92_p9
lightonocr3_4b_vllm 2 dense_tables/dense_table_018, text/text_multicolumns__2col
datalab_accurate 1 tables/SERFF_TX_random_pages 1_page1056
infinity_parser2_flash_deep 1 insurance_filing_tables/a6051a8405563b92_p9
infinity_parser2_pro_deep 1 insurance_filing_tables/a6051a8405563b92_p9
lightonocr3_0_8b_vllm 1 text/text_multicolumns__2col
muse_ember_1_0_reasoning_low_parse_with_layout_file 1 text/text_ocr__paper8col
openai_gpt_6_1_terra_reasoning_low_parse_with_layout_file 1 dense_tables/dense_table_152

Other notes

  • Cost is null for the 21 open-weight parsers LlamaIndex ran on its own hardware; leaderboard_cost_per_page_usd (the leaderboard's per-parser figure, converted from cents) is the fallback. reducto_r1_mini reports cost on only 2,780 of its pages.
  • Timestamps are UTC. A latency of 0 is a real value.
  • Atlas tiles are capped at 560 px tall to fit WebP's size limit; 6 very tall thumbnails are cropped to their top in the atlas only (listed under cropped in each atlas JSON). The thumbnails themselves are complete.
  • One page (text/text_ocr__012-25) is a 3-page PDF; its images show page 1 and page_count is 3.