hashmortar/multimodal-annual-reports

Multimodal Annual Reports — Version 3

A multimodal document question-answering benchmark with 42 corporate source documents and 1255 English questions. It includes original PDFs, reviewed HTML and Markdown, figure/table crops, reference answers, evidence and review provenance. Japanese and French sources also have English questions and answers.

MIT covers original benchmark contributions only. Publisher PDFs, reproduced text, crops, translations and other source-derived content retain third-party rights. See RIGHTS.md.

Version and scope

Version 3 adds eleven user-supplied English documents, one per GICS sector: 2,305 physical pages and 330 independently verified questions. The earlier 31 documents and 925 questions retain their canonical content. Reproduce this release with revision v3; earlier releases remain available at v1 and v2.

Item	Version 3
Accepted source documents / distinct issuers	42
Original PDF components	44
Physical source pages	6,092
Questions and reference answers	1255
Visual / table / narrative questions	358 / 490 / 407
Evidence families	1116
Indexed figure/table assets	6456
PNG files, including three supplementary Rémy Cointreau spreads	6459

The additions are ASML, AstraZeneca, BASF, Celsius Holdings, Chewy, Enel, The New York Times Company, Siemens, SM Energy, Vonovia and Voya Financial. The Chewy Q2 2026, Enel FY2025 results and SM Energy 2025 results/2026 outlook files are presentations. Celsius is a standalone Form 10-K. Actual document types and periods are retained, including Siemens' fiscal year and Chewy's quarter. Earlier sources include Netflix's 2026 proxy statement and UnitedHealth's standalone 2025 Form 10-K. The corpus therefore extends beyond annual and integrated reports.

All eleven GICS sectors are represented. Source languages comprise 37 English, four Japanese and one French document. These are user-selected corporate disclosures; country, period, length and visual richness vary substantially.

Question counts and modality mixes follow source evidence. Tables and prose are not relabeled as visuals to fill quotas. Use recorded categories when measuring visual performance. Partial Redeia remains excluded.

Loading and downloading

All four configurations use the single evaluation split. This split name does not imply a train/development/test allocation.

Configuration	Rows	Contents
qa (default)	1255	Questions, reference answers, evidence, calculations, families and review provenance
inputs	1255	Answer-free questions and document paths
documents	42	Full HTML/Markdown, source metadata and corpus paths
assets	6456	Figure/table descriptions, HTML tables and crop paths

from datasets import load_dataset
from huggingface_hub import snapshot_download
from pathlib import Path

repo_id = "hashmortar/multimodal-annual-reports"
inputs = load_dataset(repo_id, "inputs", split="evaluation", revision="v3")
documents = load_dataset(repo_id, "documents", split="evaluation", revision="v3")
assets = load_dataset(repo_id, "assets", split="evaluation", revision="v3")
qa = load_dataset(repo_id, "qa", split="evaluation", revision="v3")
snapshot = Path(snapshot_download(repo_id, repo_type="dataset", revision="v3"))

JSONL files are under data/. Resolve corpus paths from the snapshot root. Canonical reports are under corpus/data/processed/; original PDFs are under corpus/data/raw_pdfs/, raw_pdfs_v2/ and raw_pdfs_v3/. Image paths are strings, not automatically decoded image features. See the official dataset-loading guide.

inputs excludes answers, categories, evidence locations, calculations and families. Gold and review records are available elsewhere in this repository; this view does not enforce access control. For retrieval, allowlist source PDFs, HTML/Markdown, extraction and crops. Exclude QA and all review/blind-solve artifacts from model context.

Evidence and verification

Each qa row retains identity, issuer, country, sector, languages, question, answer, category, tags, visual dependency, family ID, normalized evidence, calculation/tolerance JSON, verification path, canonical QA path and the original serialized question. qa_stage distinguishes initial, expansion, batch-03 and batch-04. Variable structured fields are serialized as JSON text to preserve source-specific content.

Evidence retains source component, physical PDF page, logical report page, printed label, location/support where recorded, HTML/Markdown anchors and asset IDs/paths. Physical page numbers are one-based within the named PDF. Logical pages can span components: Terumo's sustainability component starts at logical page 35 and Chubu's financial component at 107. Siemens has restarted printed page numbering; use physical pages and the named source rather than assuming printed labels are unique.

For v3, authors inspected original sources and assembled semantic annotations into reviewed document outputs. Separate uninformed solvers received only original PDFs and questions before reference-answer disclosure. Their answers were frozen with timestamps and hashes. Distinct reviewers checked source pages, actual crops and every question. Authors repaired findings; reviewers checked changed content and affected continuations. Evidence families group exact shared tables, figures, claims and calculation inputs transitively. Earlier cohorts retain their original review and QA provenance.

These reviews used separate AI model contexts and do not constitute human validation. No inference model was benchmarked as part of construction. Historical canonical files may retain superseded pending flags or question counts; completed acceptance records govern their scope.

Packaging checks accepted hashes, original PDF bytes, ordered pages, crop inventories, evidence mappings, links, unique question IDs and complete family partitions. provenance/files.jsonl records copied-file roles and source hashes; CHECKSUMS.sha256 covers release files. Selected review JSON is retained. References to omitted scratch renders, execution checkpoints or historical snapshots are not release data paths.

Limitations and evaluation

• ASML was rebuilt during authoring before a complete initial review was frozen. Historical records retain this sequencing deviation; acceptance relies on a later full review of the current artifacts and subsequent independent checks of repairs. Celsius and Vonovia retain earlier records that named the blind solver as QA reviewer, alongside dated corrective reviews by distinct reviewers. Original records are preserved.
• Supplied files can contain source omissions, clipped content, native-text errors, inconsistent units or rounding. Such limitations are retained in document reviews rather than silently corrected in source PDFs. JPMorgan's supplied PDF omits identified printed Form 10-K pages. Equinix and Voya have broken native glyph mappings requiring source-verified recovery.
• Visual richness and document lengths vary. Report scores by document, modality, source-language tier and release cohort. Document macro averages can reduce question-density effects.
• Natural analyst questions avoid page, figure and location hints. Evaluate support, arithmetic, units and tolerances rather than exact wording alone.
• Keep evidence families together when constructing splits; prefer issuer/report holdouts for transfer evaluation. Family membership does not establish complete statistical independence.
• Extraction, answers, translations and family assignments can still contain errors. There is no prescribed train/test split, benchmark score or universal automatic judge.

Rights and attribution

Original QA annotations, metadata, and documentation are offered under MIT only to the extent the contributor holds the necessary rights. Publisher PDFs, images, reproduced text, translations and other document-derived content are not covered by that grant; their owners’ terms remain applicable.

Public availability and user-supplied files do not establish an open redistribution license. Source notices, auditor content and trademarks retain their owners' terms. Consult RIGHTS.md, provenance/rights-records.jsonl and original notices for intended use.

Prepared by Harsh Tomar, 2026. Cite this repository and the pinned revision used in an evaluation. Issuer attribution, source URLs and PDF hashes are retained in document metadata.