hashmortar/multimodal-annual-reports

Multimodal Annual Reports — Version 2

A multimodal document question-answering benchmark with 31 corporate source documents and 925 English questions. It includes reference answers, original-PDF evidence, reviewed HTML and Markdown, and figure/table crops. Japanese and French sources have English analyst-facing questions and answers.

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

Version and scope

Version 2 adds eleven user-supplied English documents, one per GICS sector: 1,583 physical pages, 1,485 crops and 330 independently verified questions. The previous 595 questions and their canonical document artifacts are unchanged. The original release remains available at revision 75c6b96325737b5e4314485fca071f64448a1e4b; use revision v2 for this release.

Item	Version 2
Accepted source documents / distinct issuers	31
Original PDF components	33
Source pages	3,787
Questions and reference answers	925
Visual / table / narrative questions	283 / 336 / 306
Evidence families	846
Indexed figure/table assets	4,550
PNG files, including three supplementary Rémy Cointreau spreads	4,553

Thirty documents have 30 questions each. Rémy Cointreau has 25, justified during review to avoid repetitive additions. Partial Redeia and its ten questions remain excluded. The broader 165-issuer plan is a future goal.

Source language	Documents	Questions	Answer language
English	26	780	English
Japanese	4	120	English
French	1	25	English

All eleven GICS sectors are represented. Seventeen issuers are from Japan, ten from the United States, one from Ireland, one from France, one from New Zealand, and one is the UK/Australia Rio Tinto group. This is a geographically concentrated, user-selected corpus.

The new documents are Caterpillar, Chevron, Costco, Duke Energy, Equinix, Home Depot, JPMorgan Chase, Linde, Microsoft, Netflix and UnitedHealth. Netflix is a 2026 proxy statement with fiscal-2025 compensation disclosures; UnitedHealth is a standalone 2025 Form 10-K. Their actual document types, reporting periods and publication dates are retained. This release therefore extends the original annual/integrated-report selection scope.

The additions contain 45 visual / 157 table / 128 narrative questions and 295 evidence families. Microsoft, UnitedHealth and Netflix have no selected visual questions: their supported questions use tables and prose, while substantive source figures remain in the extraction. Other new documents also use source-supported modality mixes. Figures or table screenshots are not relabeled to fill visual quotas. Use the recorded category when evaluating visual performance.

Configurations and loading

All four configurations use the single evaluation split. No train/development/test allocation is implied.

Configuration	Rows	Purpose
qa (default)	925	Questions, reference answers, normalized evidence, family IDs and review provenance
inputs	925	Answer-free question packets and document paths
documents	31	Full Markdown and HTML, source metadata and corpus paths
assets	4,550	Figure/table inventory, 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"
qa = load_dataset(repo_id, "qa", split="evaluation", revision="v2")
inputs = load_dataset(repo_id, "inputs", split="evaluation", revision="v2")
documents = load_dataset(repo_id, "documents", split="evaluation", revision="v2")
assets = load_dataset(repo_id, "assets", split="evaluation", revision="v2")
snapshot = Path(snapshot_download(repo_id, repo_type="dataset", revision="v2"))

JSONL files live in data/qa.jsonl, data/inputs.jsonl, data/documents.jsonl and data/assets.jsonl. Explicit configuration paths prevent unrelated canonical JSON from entering the loader. Original layout is preserved under corpus/data/processed/, corpus/data/raw_pdfs/ and corpus/data/raw_pdfs_v2/. Resolve package-relative file paths from the downloaded snapshot root. Image paths are strings, not automatically decoded Image features. See the official dataset-loading guide.

The inputs rows exclude answers, evidence locations, categories, families, calculations and gold/review content. Gold is available elsewhere in the same repository; this view does not enforce access control. For retrieval, allowlist source PDFs, HTML/Markdown, extraction content and crops. Exclude canonical QA, normalized qa, blind-solve records and all review artifacts from the model context.

Fields and source traceability

qa includes identity/context fields (id, document_id, issuer, country, gics_sector, source/answer language), question, answer, category, tags, visual_dependency, evidence_family_id, evidence, serialized calculation/tolerance fields, verification status/review path, canonical QA path, authoring stage and raw_question_json.

qa_stage is initial or expansion for historical questions and batch-03 for the 330 additions. visual_dependency is nullable text; two structured dependency records are serialized as JSON strings. Variable calculation/tolerance structures and original question objects are serialized to retain their source-specific content without incompatible column types.

Each evidence record retains source_pdf_path, source URL, one-based physical_pdf_page, logical report_page, printed label, location, support where supplied, HTML/Markdown anchors, asset IDs/paths, original-language evidence and the original serialized evidence object. Missing optional support text remains null rather than being invented. Singular and plural canonical asset references are normalized into lists.

Physical page numbers apply within the named PDF component. Logical report pages index the combined extraction. Terumo's sustainability component begins at logical page 35; Chubu's financial section begins at logical page 107. For example, Chubu logical page 110 is financial-PDF physical page 4. Use the named component instead of applying a global offset.

Canonical QA, report HTML/Markdown, extraction records, asset inventories, source PDFs, crops and selected audit records are retained. Some historical canonical verification dictionaries retain superseded pending flags or ten-question counts. Normalized acceptance follows the completed QA-expansion-01 reviews for prior additions and batch-03 reviews for new documents. Original canonical files were preserved.

Construction and verification

Authors inspected original PDFs and kept reference QA separate from document inputs. Distinct, uninformed solvers received questions and original sources before gold disclosure. Blind answers were frozen with timestamps and hashes. Separate-from-author reviewers checked all original pages, every actual crop and question support/calculations. Authors repaired findings; reviewers rechecked changed content and affected continuations. Shared-source families group exact tables, figures, claims and derivation inputs transitively.

Historical work used Astra and Sol models. The prior expansion used GPT-6.1 Sol; new workers used their configured inherited settings, as recorded in provenance. These were model contexts, including contexts from the same model family, and do not constitute human validation. No inference model was benchmarked as part of construction.

The package verifies accepted QA/review hashes, original PDF hashes, ordered pages, crop inventories, evidence/page mappings, HTML/Markdown links, unique question IDs and complete family partitions. Source-based reviews supply semantic checks beyond those mechanical tests. Every copied file's source hash and role are recorded in provenance/files.jsonl; CHECKSUMS.sha256 covers release files. Selected review JSON is included; historical references to omitted scratch files are not release data paths.

Source limitations and evaluation

• JPMorgan's supplied PDF explicitly omits printed Form 10-K pages 1–42 and 328. Every supplied physical page is retained, but this is not a claim that the file contains every page of the publisher's full report.
• Equinix's broken native glyph mappings required recovery checked against visible original pages. Garbled native text was not accepted as gold.
• Netflix's source option tables contain shifted exercise-price rows. PNGs preserve original positions; semantic alignment and its limitation were independently reviewed and documented.
• Source discrepancies, raster quality, rounding, multipart boundaries and differing source measures remain explicit in per-document reviews. Extraction, answers, translations and family assignments may still contain errors despite review.
• New documents vary substantially in visual richness. Report scores by document, modality and source-language tier; a macro average can reduce question-density effects.
• Natural analyst questions generally avoid page, chart, color and location hints. Compare source support, arithmetic, units and tolerances rather than requiring verbatim answers.
• If creating training/validation splits, keep evidence families together and prefer issuer/report holdouts for transfer evaluation. Families do not establish complete statistical independence.
• This release has no prescribed split, benchmark scores, universal automatic judge or claim of general financial-analysis competence.

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.

Issuer PDFs, reproduced report text, figures, tables, translations, source crops, auditor content and trademarks retain their owners' terms. Public access or user-supplied files do not establish an open redistribution license. Existing assessments include local/personal-use and third-party restrictions; this publication does not grant additional rights or establish satisfaction of those terms. Consult RIGHTS.md, provenance/rights-records.jsonl and source notices for intended use.

Prepared by Harsh Tomar, 2026. Cite the repository and pinned revision used in an evaluation. Source URLs, issuer attribution and PDF hashes are retained in document metadata. Version 1 remains at its original revision; version 2 appends the accepted sources and QA without rewriting the earlier canonical objects.