1. Overview

This dataset provides longitudinal financial data — balance sheet items, profit and loss aggregates, employee counts, and CIC34 community-impact narratives — for nonprofit companies registered at Companies House. It was produced by the UK Third and Civil Society Sector Database project, which collects, processes, and links public administrative data on civil society organisations across the United Kingdom.

The data is drawn from the annual accounts that nonprofit companies file at Companies House. Two complementary sources are used: structured XBRL filings (machine-readable accounts) and PDF accounts processed through a structured-output LLM extraction pipeline. The merged dataset gives broader coverage than either source alone, particularly for smaller filings such as Community Interest Company (CIC) abridged accounts.

The dataset spans financial years from 1981 to 2026 and contains 1,342,198 financial-year records linked to 217,352 nonprofit organisations. Each row represents one company’s accounts for a single financial year.

2. What are Nonprofit Financial Records?

Companies registered at Companies House are required to file annual accounts. The level of detail filed varies by company size — micro-entities and small companies file abridged or filleted accounts, while medium and large companies file full statutory accounts. CICs additionally file a CIC34 community-impact report alongside their accounts.

XBRL Accounts (Companies House monthly extracts)

Companies House publishes monthly bulk extracts of accounts filed in machine-readable XBRL format. The extracts contain structured tagged values for balance sheet items, profit and loss aggregates, and employee counts. These are downloaded from the Companies House accounts download site.

PDF Accounts (LLM-extracted)

For older filings and CIC abridged accounts that pre-date or were filed outside the XBRL stream, the project extracts financial line items from the original PDF accounts using a structured-output LLM pipeline. The extracted values are mapped onto the same column schema used for XBRL, so the merged dataset has a uniform shape regardless of source.

CIC34 Community Interest Reports

Community Interest Companies file a CIC34 alongside their accounts. The CIC34 contains four narrative sections: the company’s activities and impact, stakeholder consultation, directors’ remuneration, and any asset transfers. These narratives are extracted from the same PDF filings.

3. Dataset Contents

The dataset is organised into seven column groups. Each row represents one company-year record. Where a value is missing, the cell is left blank.

Identifiers and Time Period

Aggregate Financials

Balance Sheet

Balance-sheet line items including tangible_fixed_assets, debtors, cash_bank_in_hand, current_assets, creditors_due_within_one_year, creditors_due_after_one_year, net_current_assets_liabilities, total_assets_less_current_liabilities, net_assets_liabilities_including_pension_asset_liability, called_up_share_capital, profit_loss_account_reserve, and shareholder_funds.

Profit & Loss Detail

Detailed P&L items including cost_sales, gross_profit_loss, administrative_expenses, raw_materials_consumables, staff_costs, depreciation_other_amounts_written_off_tangible_intangible_fixed_assets, other_operating_charges_format2, profit_loss_on_ordinary_activities_before_tax, tax_on_profit_or_loss_on_ordinary_activities, wages_salaries, dividends_paid, and government_grant_income.

Staff

average_number_employees_during_period: average number of persons employed during the financial year.

CIC34 Narratives (CICs only)

Four free-text fields extracted from the CIC34 form: cic34_activities_impact, cic34_stakeholder_consultation, cic34_directors_remuneration, cic34_asset_transfer.

Metadata

Provenance and lineage columns including source_file, file_type, taxonomy, balance_sheet_date, from_prior_year (whether the row came from a prior-year companion filing), source_dataset (xbrl or pdf_extraction), is_cic, and csotype (CIC / Charity / Co-operative / Mutual / Other).

4. Coverage & Completeness

The combined dataset draws from XBRL and PDF sources. Coverage varies by financial year, by company size, and by individual line item. The charts and tables below summarise where the data is fullest and where it is sparsest.

Coverage is reported as a diagnostic description of this release. It is not a guarantee that every filing or line item is present, and the release policy does not impose arbitrary absolute coverage floors.

Records by Source and Year

Figure: Records by source and year. Earlier years are XBRL-dominated; recent CIC filings come predominantly via the PDF extraction stream.

Source Overlap

Figure: Most records are uniquely sourced; the XBRL stream is preferred where both are available.

Variable Availability

Figure: Fill rates for the financial columns across all records.

Headline Coverage

Year-by-Year

5. What Can You Learn?

The dataset supports questions about the size, financial health, and activity of the UK nonprofit company sector. A non-exhaustive list of uses:

Sector Growth Trends

Figure: CICs with reported accounts and median net assets by financial year.

Financial Profile

Figure: Distribution of five high-coverage balance-sheet fields across the sector.

Employment Dynamics

Figure: Total employees and reporting organisations by financial year for CIC accounts.

CIC34 Narratives

The CIC34 narrative fields support qualitative analysis of community impact and stakeholder engagement. Coverage rates by field:

Director Remuneration

Director remuneration patterns can be inferred from the CIC34 directors’ remuneration narrative. Among CICs reporting remuneration: coverage rate 84%, median statement length 33 characters.

6. Limitations & Caveats

Reporting Lag

Accounts are typically filed 6–12 months after the financial year ends. The most recent financial years will therefore be incomplete.

Filing Thresholds

Many CICs file as micro-entities and are exempt from disclosing detailed P&L information. Coverage of fields such as turnover_gross_operating_revenue and staff_costs is materially lower than balance-sheet coverage as a result.

Variable Financial Year Periods

Financial years vary in length, particularly for newly incorporated or dormant companies. Compare values per-year cautiously when the FY length differs from twelve months.

Empty PDF Extractions

The conversion quarantined 10,717 source filings whose extraction contained no usable values: 10,243 were not classified as dormant and 474 were classified as dormant. 1 local source PDF was unavailable for review. These cases remain in the release evidence rather than being filled with inferred values.

All-Zero PDF Rows

Among current PDF rows that are not explicitly dormant (including rows where dormancy is unknown), 8,233 of 124,675 have all twelve balance-sheet fields recorded as zero or blank (6.604%). This exceeds the unchanged 2.000% default diagnostic threshold. It is an accepted release exception and an extraction pattern, not evidence that every underlying company had a zero balance sheet.

Conflicting PDF Candidates

The converter recorded 81,734 material conflict events, representing 46,048 distinct source-period-field cells omitted from the converted PDF data. It emitted 0 disputed values. When direct candidates disagree, no winner is selected: the published cell is blank unless a non-conflicting preferred source supplies it. In the assembled final population, 28,724 affected cells remain blank and 10,792 are safely populated by a separate, non-conflicting auditable source.

Disabled Arithmetic Derivations

The pipeline does not manufacture additional coverage using the following identities because their PDF-source confirmation checks failed:

• creditors_due_within_one_year = current_assets - net_current_assets_liabilities — PDF identity check failed.
• current_assets = net_current_assets_liabilities + creditors_due_within_one_year — PDF identity check failed.

How to Interpret Blanks

A blank can mean that the filing did not disclose the item, extraction produced no usable value, or conflicting candidates were deliberately suppressed. Treat blanks as missing information, not as zero. The source_dataset column can be used to distinguish XBRL and PDF provenance.

Spine Filtering

Only nonprofit companies present in the project Spine (the deduplicated register of UK third-sector and civil-society organisations) are included. Profit-making companies and other Companies House registrants are excluded.

What is NOT in the Data

The dataset does not include cash-flow statements, notes to the accounts in full, contingent liabilities, related-party transactions, or auditor qualifications. For these, consult the original PDFs at Companies House.

7. Citation & Licence

The dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) licence. You are free to share and adapt the data for any purpose, including commercially, provided you give appropriate credit.

Suggested Citation

McDonnell, D. et al. (2026). Nonprofit Financial Records.
UK Third and Civil Society Sector Database.
Available at: https://uk-third-sector-database.github.io/data/
Licensed under CC BY 4.0.

8. Changelog

v1.2 — July 2026

• Rebuilt the XBRL and converted-PDF inputs and refreshed all figures and tables from the exact release candidate.
• Applied source-backed scalar/date/employee corrections and a pinned mapping no-regression audit.
• Removed arbitrary default coverage floors; coverage remains a diagnostic.
• Documented accepted empty-extraction, all-zero-row, and suppressed conflict limitations from the passing release manifest.
• Kept both unconfirmed arithmetic coverage derivations disabled.

v1.1 — May 2026

• Incorporated ~42,000 newly extracted CIC PDF accounts.
• Added csotype column (CIC / Charity / Co-operative / Mutual / Other).
• Refreshed coverage statistics throughout the guidance.

v1.0 — February 2026

• Initial public release of the merged XBRL + PDF combined dataset.

A1. Pipeline Architecture

The dataset is produced by a multi-stage pipeline that combines two source streams into a single deduplicated CSV.

Stream 1: XBRL Accounts (monthly bulk extracts)

Companies House publishes monthly ZIP archives of XBRL-tagged accounts. Each archive is downloaded, parsed using stream-read-xbrl, and appended to a per-month CSV with a uniform schema. The combined extract represents the ‘XBRL’ stream.

Stream 2: PDF Accounts (LLM-extracted)

For each PDF account filing, the pipeline (a) renders pages to JPEG images, (b) submits the image set to OpenAI’s Batch API with a JSON-schema response format that prompts a structured extraction of line items, balance-sheet date, and CIC34 narratives, and (c) collects the JSON outputs to disk under data/output/accounts-extractions/json-schema/{coyno}/{txn}.json.

Stream Convergence: XBRL Format Conversion

The PDF JSON extractions are converted into the same column schema as the XBRL stream using an approved label mapping plus a Jaccard-similarity fallback for unmapped labels. Sign-aware columns (creditors, expenses) are normalised to positive magnitudes.

Merge and Provenance

The XBRL and PDF-converted CSVs are merged on (coyno, fy). XBRL is preferred where both sources cover the same record. A source_dataset column records each row’s provenance.

A2. Source Data

Companies House — XBRL Accounts

The official monthly bulk download of accounts in XBRL format is available at download.companieshouse.gov.uk/en_monthlyaccountsdata.html. Historic archives are at historicmonthlyaccountsdata.html.

Companies House — PDF Accounts

For filings outside the XBRL stream (older accounts, CIC abridged filings, etc.), original PDFs are obtained via the Companies House document API and stored locally for extraction.

Companies House — CIC34 Forms

Community Interest Reports are filed as part of the same PDF accounts package for CICs and are extracted alongside the financial line items.

Spine

The hash-bound release Spine (code/release-v1.2-Jul2026/TSCS_spine.spine.csv) is the deduplicated register of UK third-sector and civil-society organisations. It is the inclusion list for the dataset and supplies uid, normalised_name, is_cic, and the derived csotype.

A3. Deduplication & Validation

Within-Source Deduplication

For each (coyno, fy) pair, the pipeline prefers current-year filings over prior-year companion rows. Same-source duplicate rows (for example, balance-sheet and P&L sections of the same filing) may contribute non-conflicting values to one row. It does not simply take the first non-null value: distinct direct candidates for the same source-period-field cell are treated as a material conflict, logged, and omitted. This release logged 81,734 such events and omitted 46,048 disputed cells and emitted 0 unsafe disputed values. In the assembled final population, 28,724 of the affected cells remain blank and 10,792 are populated from a separate, non-conflicting auditable source.

Cross-Source Deduplication

Where both XBRL and PDF sources cover the same (coyno, fy), the XBRL row is preferred. The PDF row is dropped and not retained as a duplicate.

Date Validation

fy is normalised to integer-string. Where fy is blank but balance_sheet_date is populated, fy is derived (with a one-year offset for prior-year companion rows). Missing fye is imputed from the company’s most common financial-year-end month-day.

Sign Normalisation

XBRL stores expenses and liabilities as positive magnitudes; PDF extractions sometimes preserve negative accounting signs. The pipeline applies abs() to nine sign-aware columns (creditors_*, cost_sales, administrative_expenses, etc.) so both sources use the same convention.

Arithmetic Derivations

Potential balance-sheet identities are disabled unless they pass source-stratified confirmation checks. The two candidate derivations in this release remain disabled because their PDF identity checks failed.

Employee-Count Guardrails

Legitimate non-negative fractional employee counts are preserved. Values tagged as currency or percentages, negative values, and counts above 100,000 are rejected. The same guard applies to direct, recovered, and prior-year candidates; if a current employee fact is rejected, the pipeline does not silently replace it with a prior-period value.

Spine Filtering

Only company numbers present in the Spine are retained. A left-join on uid populates is_cic and the derived csotype.

A4. Reproducibility

Code

The end-to-end pipeline lives in code/companies-house/:

• pdf_accounts_extraction/openai-api/ — PDF-to-JSON extraction (rendering, batch submission, batch processing).
• pdf_accounts_extraction/xbrl_format_conversion/ — JSON-to-XBRL column schema conversion, label mapping, dataset merge, and public-zip publishing.
• reporting/guidance/ — this guidance document generator.

Dependencies

Managed via the uv package manager. Run python code/companies-house/pdf_accounts_extraction/xbrl_format_conversion/dependencies.py to install conversion-pipeline dependencies; analogous scripts cover the other modules.

Running the Pipeline

The root ordinary build now names its two formerly hidden final stages. convert-pdf creates one explicit, versioned PDF CSV and merge-final consumes that exact path to build the actual final financial dataset. combine-full remains the XBRL-only intermediate; release gating and publication remain separate.

cd code/companies-house
python run_pipeline.py convert-pdf --xbrl-csv XBRL.csv --output CONVERTED_PDF.csv --diagnostics-dir PDF_DIAGNOSTICS
python run_pipeline.py merge-final --xbrl-csv XBRL.csv --pdf-csv CONVERTED_PDF.csv --spine-csv SPINE.csv --matches-csv MATCHES.csv --output COMBINED.csv

cd pdf_accounts_extraction/xbrl_format_conversion
python run_pipeline.py release 2026-07-22-spine-v1.2-final-r3 --converted-csv CONVERTED_PDF.csv --conversion-report CONVERSION_REPORT.csv --check-b-audit CHECK_B_AUDIT.json --xbrl-csv XBRL.csv --spine-csv SPINE.csv --matches-csv MATCHES.csv --output-dir OUTPUT_DIR --accepted-limitations-policy POLICY.json

python ../../reporting/guidance/generate_guidance.py --combined-csv cso-spine-nonprofit-company-accounts-combined-2026-07-22-spine-v1.2-final-r3.csv --pdf-csv cic-accounts-extractions-xbrl-format-2026-07-22-spine-v1.2-final-r3.csv --release-manifest manifest.json --release-date 2026-07-22 --release-version v1.2

python run_pipeline.py finalize-release --manifest manifest.json --guidance-pdf GUIDANCE.pdf --guidance-html GUIDANCE.html --licence-template LICENCE.txt --release-repo RELEASE_REPO --website-guidance WEBSITE_GUIDANCE --public-zip-name PUBLIC.zip --archive-date 2026-07-22

Guidance generation verifies the SHA-256 of both CSV inputs against the passing candidate manifest. Finalization does not rebuild that immutable candidate: it verifies the candidate and retained limitation evidence, then binds the explicit guidance pair, licence, and destinations by SHA-256 in a separate publication manifest before promotion. A lexical “latest file” lookup is not part of the release procedure.

Source Code

The release repository is at github.com/uk-third-sector-database/tso-database-builder.