README.md

---
pretty_name: FilingsBench
license: apache-2.0
task_categories:
- table-question-answering
language:
- en
tags:
- sec-filings
- financial-document-analysis
- inline-xbrl
- information-extraction
size_categories:
- n<1K
---

# FilingsBench

An accession-linked, issuer-disjoint benchmark for auditable extraction from SEC 10-K, 10-Q, and
8-K filings. The financial-facts v0.1 slice contains 60 filings from 15 issuers and eight fields;
its labels are explicitly **silver SEC cross-source**, not human-gold. Narrative slices use
`candidate_pending_independent_review` until blinded review and adjudication are complete.

The repository contains provenance and labels, not copies of filing HTML. See
[FinStruct](https://github.com/arnavprabhu/uraion-finstruct) for builders, raw predictions,
metrics, schemas, and known failures.

Financial-facts dataset SHA-256:
`fbac61331b715ac196e6af57dc8276c40d3819b31489cf77c7a7914bb17b5725`.

## Controls v0.1 candidate slice

The `controls/` directory adds 12 issuer-disjoint filings and four narrative fields: ICFR
effectiveness, disclosure-controls effectiveness, material-weakness presence, and remediation
status. The deterministic v0.2.0 baseline matched all 48 candidate outcomes and attached exact
evidence to every present prediction. Labels remain `candidate_pending_independent_review`.

Controls dataset SHA-256:
`9de0c23ed5e971d9426b7d235b52332fb13c257829d4a09743641eef213717de`.

[Controls report](https://github.com/arnavprabhu/uraion-finstruct/blob/main/docs/controls_benchmark_report_v0.1.md)

## Form 8-K Item 4.02 v0.1 candidate slice

The `non_reliance/` directory adds 14 issuer-disjoint Form 8-K primary documents and six fields:
Item 4.02 paragraph variant, non-reliance scope, latest conclusion/notice date, restatement status,
auditor-discussion disclosure, and controlled accounting topics. Twelve records contain Item 4.02;
two are negative 8-K controls that should abstain on every field.

The deterministic v0.3.0 baseline matched all 84 candidate outcomes, attached evidence to every
present prediction, and passed 82/82 normalized-document offset checks. The first run exposed an
incomplete candidate label for one filing; exact evidence corrected the label without a parser
change. These labels are not human-gold or a prevalence sample.

Item 4.02 dataset SHA-256:
`2113ed77793a52c133d674c74c29238d684e77c3fdff1888f92d97dc5b9b594e`.

The directory includes a label- and prediction-free `review_packet.jsonl` for blinded independent
review, with SHA-256
`3ea7169169698ea9155df7d90705719df46fe3e799a37e5814911643ec814aec`.

The v0.3.1 reviewer kit adds `review_template.csv`, a spreadsheet handoff generated from the blind
packet, plus its submission schema and walkthrough. FinStruct validates all 14 records and 84
field decisions, enforces controlled values and abstention rationales, and can verify exact quotes
against SEC filed text before emitting normalized JSONL offsets. The template SHA-256 is
`ca24011b90c85204d581aefeca8a5cdbee3a239fff1cab5aed2c570540c5d027`.

The v0.3.3 starter kit adds `review_packet_starter.jsonl` and
`review_template_starter.csv`: four filings and 24 field decisions selected only by ascending
`sha256(starter-v0.1:record_id)`. The packet remains label- and prediction-free and uses the same
validator as the full review. A validated starter is partial independent coverage, not a complete
slice review or human-gold claim. The packet SHA-256 is
`17954faa8388889b6ca59ecec40be39a377bd01c671e43ae5251947397299cab`; the template SHA-256 is
`22a9d97c1dc68559e93520fe134706750aa4f4505236baff882c13701622f8e9`.

FinStruct v0.5.0 adds deterministic comparison and adjudication for two normalized submissions.
It isolates status, value, and exact-evidence disagreements; requires a third distinct adjudicator;
and re-verifies every present source and final quote against SEC filed text before emitting the
`human_gold_double_review_adjudicated` state. This is promotion infrastructure only: the dataset
currently contains zero independent submissions and zero human-gold records.

[Item 4.02 report](https://github.com/arnavprabhu/uraion-finstruct/blob/main/docs/non_reliance_benchmark_report_v0.1.md) ·
[Independent review walkthrough](https://github.com/arnavprabhu/uraion-finstruct/blob/main/docs/independent_review_item_402.md) ·
[Tutorial](https://github.com/arnavprabhu/uraion-finstruct/blob/main/docs/tutorial_item_402.md) ·
[v0.5.0 release](https://github.com/arnavprabhu/uraion-finstruct/releases/tag/v0.5.0)

## Zero-setup v0.3.2 quickstart

[Open the executed notebook in Colab](https://colab.research.google.com/github/arnavprabhu/uraion-finstruct/blob/main/notebooks/finstruct_item_402_quickstart.ipynb)
to install the immutable v0.3.2 wheel, verify one tagged temporal-audit filing hash, extract the six
Item 4.02 fields, and round-trip all seven emitted evidence offsets. It requires no repository
checkout, model, GPU, inference API, SEC credential, or private document.

The notebook reproduces installation and system behavior on unlabeled development data. It does
not establish accuracy, human-gold status, post-tuning generalization, or production quality.

## Item 4.02 temporal audit v0.2

The `temporal/` directory adds a label-free 25-issuer audit frozen before prediction from a fixed
May 1–July 27, 2026 SEC EFTS query. It contains the exact selection manifest, accession-linked
records, and separate v0.3.1 baseline and v0.3.2 development-replay outputs. The records SHA-256
is `0044db83e2782d4a02f75f7b9f376274a0ea6f056129e22fb8eec441069e620e`.

Accuracy is explicitly null. Maintainer inspection changed 37 field outputs across 18 filings;
the final development replay emitted 141/150 fields and passed 154/154 evidence-offset checks.
Because the set was used for development, it is not human-gold, an untouched post-tuning holdout,
market prevalence, or production accuracy.

[Temporal audit](https://github.com/arnavprabhu/uraion-finstruct/blob/main/docs/non_reliance_temporal_audit_v0.2.md)