damo-da/oag-nepal-audit-reports

OAG Nepal Audit Reports — Nepali transcripts and ruled tables

Machine-readable transcripts of 6,235 publications of the Office of the Auditor General of Nepal (महालेखा परीक्षकको कार्यालय, OAG) — the annual audit reports of local governments, provinces and central bodies, plus the OAG's own bulletins, journals and financial statements.

The OAG publishes these as PDFs whose text layer is, for most documents, legacy pre-Unicode Devanagari: fonts like Preeti and Fontasy Himali that map Nepali glyphs onto Latin code points, so an ordinary extractor returns Latin-looking gibberish. This dataset is the decoded, page-anchored text, plus the ruled audit tables as a cell grid — the itemised findings, voucher numbers and rupee amounts that are the substance of an audit report.

335,392 pages. 558,081,550 Devanagari characters. 444,139 ruled tables over 30,271,523 cells, 5,680,242 of them parsed numerics.

What is in it

	rows	what
documents	6,235	one row per publication, with source URL, SHA-256 of the PDF and of the transcript, and page/table rollups
pages	335,392	page-by-page markdown, one row per source page, with per-page provenance
tables	444,139	one row per ruled table block, with its introducing sentence and shape
table_cells	30,271,523	long form, one row per (row, col) slot, with parsed numeric values
quality	6,235	8 audit axes per document, with the underlying counts, not just a verdict
figures	181,942	one row per raster image placed on a page, classified — see Figures below, because most of these are not figures

Plus markdown/ — the 6,235 transcripts themselves, which are the primary artifact — pdf/ — the 6,236 source PDFs they were made from, 19.36 GiB — and MANIFEST.json, listing every file with its size and SHA-256.

By publication type

bucket	documents
local-level-report	6,007
report_province-report	56
publication_audit-bulletin	29
publication_right-to-information	22
publication_audit-journals	21
report_annual-report	18
report_performance-audit-report	17
publication_annual-progress-report	14
report_report-summary	13
report_state-owned-enterprises-report	10
report_it-audit-report	9
publication_financial-statement	6
report_special-audit-report	5
publication_auditor-general's--work-achievement	3
report_environmental-audit-report	2
publication_sai-performance-report	2
publication_other	1

Local-level reports by province

province	reports
Koshi Province	1,101
Madhesh Province	1,067
Bagmati Province	951
Lumbini Province	873
Sudurpashchim Province	702
Gandaki Province	681
Karnali Province	632

Most-represented fiscal years (BS)

fiscal year	documents
2077	796
2075	776
2078	775
2082 (2080/81)	765
2081 (2079/80)	752
2079 (2078/79)	752

🛑 How to read the quality of this data

This is a decoded legacy-font corpus, not a clean digital text. Read these five limits before using it for anything that depends on a number being right.

1. Accuracy is an AGREEMENT rate against a second instrument, not a truth claim

The corpus was measured against an independent vision-OCR read of 390 pages sampled across quality strata — a reader that never touches the text layer, so it fails differently from font-table decoding. Corpus-weighted word agreement 76.2%. On the 171 of 390 pages where both instruments produced comparable amounts of text, clean-verdict pages agree at about 93%.

⚠️ Neither figure is "accuracy". Where the two instruments agree both are almost certainly right; where they disagree one is wrong and the pair only localises the error. The gap between 76% and 93% is mostly the corpus over-producing text, which is the defect described next.

🛑 Those figures were measured on markdown-quality-v16, NOT on this build's tree (markdown-quality-v17-d22f13e13bfe9d0d), and have not been re-measured since. They therefore predate whatever this build changed, and the de-duplication described below should move them upward — the same measurement predicted clean agreement near 93% once over-production was removed. Read 76.2% as a floor established on an earlier tree, not as this build's measured agreement.

2. Some pages still contain their own content twice

364 of 335,392 pages (0.109%) contain their own content twice — down from 2,423 in the v16 release tree this supersedes. Duplicated Devanagari is well-formed Devanagari, so no character-level quality axis can see it — use the structural test (is the page's token stream exactly its own first half twice?) if it matters to you.

The residual is not unexplained. On 198 of 200 examined, a coarse detected table holds figures the finer grid lost, so de-duplicating would delete real rupee amounts. Leaving the page doubled is the cheaper error.

3. Digits are never folded across scripts, and you should not fold them either

The source typesetting mixes scripts: this build carries 53,404,486 Devanagari digits and 16,681,948 ASCII digits. २०७९ and 2079 are kept distinct everywhere, text is always verbatim, and value_num is filled only where a cell parses cleanly as a scalar — a fiscal year (२०७९।८०), a voucher number with a trailing danda, or a range is left null rather than guessed.

4. Per-page provenance: not every page came from the PDF's text layer

pages.source	pages	what it means
text_layer	325,764	decoded from the PDF's own text layer by likhit
vision_ocr	9,628	the PDF page had under 80 characters of text layer, so the page was rendered at 300 DPI and read by a vision model

⚠️ A vision read substitutes whole consonants invisibly — a wrong character is still valid Devanagari, with no replacement character and no ratio change to give it away. Treat vision_ocr pages as the lower-confidence arm.

5. A few documents were never decoded at all, and you must be able to see which

31 documents contain zero Devanagari characters, and 13 of them score clean. That is not a contradiction and it is the clearest illustration on this card of what a verdict is worth: the corpus holds two entirely different populations that both come out with no Devanagari.

• Most are English by design — OAG publishes unofficial English translations of its annual report summaries, and those are correct, complete and rightly clean.
• 14 are documents where likhit produced no usable text, so the markitdown fallback shipped the undecoded legacy bytes: text that looks like v n s T i s rather than Nepali. The audit catches these — they fail legacy_ascii and structure — but documents.engine reads markitdown+likhit for them, the same value as a document likhit decoded successfully.

So documents.engine_fallback (bool) and engine_fallback_reason are shipped to separate them. Filter on engine_fallback == False if you want text that was actually decoded.

🛑 These are not spread evenly, and in the worst-scoring collections they are the whole story:

collection	documents	fallbacks	of the bucket's garbled	the rest
publication_auditor-general's--work-achievement	3	2	0 of 0	1 of 1 clean
publication_right-to-information	22	11	10 of 10	10 of 11 clean
local-level-report	6,007	1	1 of 99	5,667 of 6,006 clean

Read the publication_right-to-information row as: the collection is not hard to transcribe — 11 of its 22 documents were never transcribed, and those are 10 of its 10 garbled rows. Set them aside and the remaining 11 score 10 clean, like the rest of the corpus.

The audit verdicts, as shipped

verdict	documents
clean	5,854
suspect	269
garbled	112

🛑 A clean verdict means no axis fired, which is weaker than "correct". The 8 axes that ran are legacy_ascii, matra_damage, mojibake, numeric_damage, repetition, repha_loss, spacing, structure — every one an internal well-formedness check on damaged conjuncts, replacement characters, spacing or numeric shape. None of them compares a transcript against the page it came from, so a systematically wrong but well-formed decode scores clean.

This corpus has shipped exactly that: a known malformed conjunct-ra family runs to thousands of occurrences across hundreds of documents while each document's own rate stays under every threshold, so the axis reports it and the verdict dilutes it away. The quality frame therefore ships the axis counts in *_detail_json, not only the verdicts, so you can set your own threshold instead of inheriting ours.

The source materials

pdf/ carries all 6,236 source PDFs, 19.36 GiB, laid out as pdf/<bucket>/<file_id>__<name>.pdf. Each documents row points at its own with pdf_relpath, so a transcript and the paper it came from are one join apart.

Every staged PDF was checked byte-for-byte against the documents row that describes it — matched on documents.pdf_sha256 — so these are the exact bytes the transcripts were made from, not a re-download that might differ. source_url is also carried per row if you want the publisher's copy: 6,235 of 6,235 rows have a pdf_sha256 to check it against.

⚠️ The layout uses the file_id prefix rather than the bare file name because 229 of the documents share a file name with another document. A name-keyed layout silently overwrites them.

⚠️ pdf/ holds 6,236 files but documents has 6,235 rows. The difference is 1 PDF with no transcript: publication_financial-statement/11356__BVm-Financial Report 2075-76.pdf. Shipped anyway, and named here rather than quietly dropped — it is a scan with no text layer that no paid vision read has covered, so there is nothing to transcribe from.

Figures — and why the number is 15,585 and not 181,942

figures has 181,942 rows, one per raster image placed on a page. Most of them are not figures. The frame says which is which instead of letting the row count imply an answer:

kind	rows	what it is
inline_image	15,585	a figure candidate — a photograph, diagram or chart printed on the page
page_scan	11,082	a scan of the whole page, i.e. an image of a page this dataset already transcribes
template_furniture	66,980	an office seal, emblem or rule whose identical bytes recur across many documents
glyph_or_fragment	88,295	sub-visible vector fill or a rasterised glyph — median area around 150 pixels

kind is derived, and the thresholds are conventions — page_frac >= 0.6 is a page scan, identical bytes in >= 20 documents is furniture, and below page_frac 0.005 is a fragment. Both inputs ship per row (page_frac, n_documents_sharing_image), so re-cut them if your use needs a different split. Neither has a clean gap behind it: page_frac is bimodal but its trough is populated, and the sharing count has no gap at all.

🛑 The fragment threshold is the one that changes the answer rather than trimming it. Without it 103,880 placements read as inline images — but that class had a median page fraction of 0.0, a median area of 153 pixels, and two documents supplying 76.8% of it. Requiring half a percent of the page leaves 15,585 candidates in 3,995 documents with the largest single document down to 1.6%.

What is deliberately not here

• figure_series_data — no chart was read; CIAA's figure_data came from a paid vision read and this corpus has had none
• per_page_markdown_files — 335,392 loose files; the identical text is in pages.parquet
• extracted_image_files — 181,942 raster placements, mostly repeated seals and full-page scans; the source PDFs are shipped instead

A page anchor is a BLOCK START, not a page boundary

Inside a long ruled table the extractor emits the whole run under the anchor of the page where the table began. So a page's own section can be short while its text sits under an earlier anchor, and pages.chars / pages.n_table_blocks are per section. pages.printed_page_no carries the page number printed in the publisher's own footer where the page has one, which is the reliable way to cite a page.

Duplicates: the same report is sometimes published twice

documents.duplicate_role	rows
unique	6,137
canonical	39
unindexed	39
conflict	20

work_id groups the rows that are the same underlying work, so deduplicate on work_id rather than on doc_id if you are counting reports. republication marks a document republished under a second file id; attribution_conflict marks rows that must not be collapsed, because two municipalities each claim the report and the conflict is the finding.

⚠️ This grouping was derived on the index named in Provenance below, not re-derived on this build's tree. Which documents are the same work is a property of the source PDFs and transfers cleanly; the attribution_conflict rows were separated on content and have not been re-checked here.

Loading it

from datasets import load_dataset

pages = load_dataset("damo-da/oag-nepal-audit-reports", "pages", split="train")
cells = load_dataset("damo-da/oag-nepal-audit-reports", "table_cells", split="train")

# every numeric cell of one province's reports
docs = load_dataset("damo-da/oag-nepal-audit-reports", "documents", split="train")
karnali = {d["doc_id"] for d in docs if d["province"] == "Karnali Province"}
amounts = [c for c in cells if c["doc_id"] in karnali and c["is_numeric"]]

Provenance

version	1.2
build date	2026-08-17
transcript tree	markdown-quality-v17-d22f13e13bfe9d0d
extractor revision	2822cdd (likhit)
build run	d22f13e13bfe9d0d
audit record	runs/vol744-fix-d22f13e13bfe9d0d/audit-v17-d22f13e13bfe9d0d.json
duplicate grouping	corpus-v16.sqlite
dataset size	21.80 GiB over 12,479 files
characters (with page anchors)	819,997,018
Devanagari characters	558,081,550 (68.1%)

Transcription is by likhit, an open-source Nepali PDF text extractor, at the revision named above.

Licence and intended use

The underlying documents are public records published by the Office of the Auditor General of Nepal. This derived dataset is released under CC BY-NC 4.0 for research, journalism and civic accountability work.

It is a research artifact, not an official record. A figure that matters should be checked against the source PDF, which source_url and pdf_sha256 let you do. Do not present a transcript as the OAG's own text.