damo-da/ciaa-annual-reports

CIAA Annual Reports — Nepali transcripts, ruled tables and chart data

Machine-readable transcripts of the annual reports of Nepal's Commission for the Investigation of Abuse of Authority (अख्तियार दुरुपयोग अनुसन्धान आयोग, CIAA) — all 35 it has published to date. The 1st to 35th reports, fiscal years BS 2047/48 – 2081/82 (AD 1990–2025).

The CIAA publishes these as PDFs whose text layer is, for several years, legacy pre-Unicode Devanagari that ordinary extractors turn into Latin-looking gibberish. This dataset is the repaired, page-anchored text, plus the two things the text layer does not contain: the ruled statistical tables as a cell grid, and the charts — bar/pie/line figures whose numbers exist only as pixels — read off 300-DPI page renders.

Nepal's anti-corruption record is documented almost entirely in the CIAA's annual report series. That is a claim about the series, not about this dataset: what is here is the 35 reports listed below, and the coverage caveats further down say what within them is and is not machine-readable.

What is in it

	rows	what
documents	35	one row per report, with source URL, SHA-256 of the PDF and transcript, and page/quality rollups
pages	10,015	page-by-page markdown — one row per source page, with per-page provenance
tables	1,735	one row per ruled table, with its introducing sentence and shape
table_cells	678,165	dense cell grid (262,667 non-empty), with numerals parsed across both digit scripts
figures	477	one row per chart, with type, axes, unit and verification note
figure_data	6,116	every chart data point: label, series, value, and whether it was read off the axis
quality	35	per-report quality axes and findings, so the caveats travel with the data

And, as files:

path	files	size
figure_images/	526	400.6 MB
pdf/	35	341.9 MB
ciaa_corpus.sqlite + card	1	165.1 MB
markdown/	35	67.5 MB
pages/	10,015	67.0 MB
data/	7	18.8 MB

pages/<fy>/pNNNN.md is the page-by-page transcript as loose markdown; markdown/ holds the full anchored transcript per report; pdf/ holds the unmodified source PDFs so any value can be checked against the original; figure_images/ holds the 300-DPI renders the chart data was read from. ciaa_corpus.sqlite carries every table above, indexed, plus three views (v_page_citable, v_table_numbers, v_figure_series).

Quick start

from datasets import load_dataset

pages = load_dataset("damo-da/ciaa-annual-reports", "pages", split="train")
print(pages[0]["text"][:400])

# Chart data points, already numeric
figs = load_dataset("damo-da/ciaa-annual-reports", "figure_data", split="train")

Or query the SQLite directly — usually the fastest route for the tables:

import sqlite3, pandas as pd
from huggingface_hub import hf_hub_download

db = hf_hub_download("damo-da/ciaa-annual-reports", "ciaa_corpus.sqlite", repo_type="dataset")
con = sqlite3.connect(db)

# every chart series in one frame
pd.read_sql("SELECT * FROM v_figure_series", con)

# complaint counts from the ruled tables, only where the page citation is exact
pd.read_sql('''
  SELECT doc_id, page_no, intro_text, value_num
  FROM v_table_numbers
  WHERE anchor_fidelity = 'tight' AND intro_text LIKE '%उजुरी%'
''', con)

Quality — measured, not asserted

Every figure below was measured on the bytes in this repository, by the build that produced them. That build refuses to emit a dataset that fails any of its 14 assertion groups — page-row count against each PDF's own page count, contiguous page numbering, dense-grid integrity, referential integrity across tables, post-copy SHA-256 of every artifact, and the BS→AD mapping among them.

Text. 27,049,673 characters, 19,946,598 of them Devanagari. 0 U+FFFD replacement characters. 0 control characters. 14 private-use characters remain — every one of them a recorded unmapped CID, which is a deliberate marker meaning "the source PDF's font gave no way to read this glyph", kept visible rather than dropped so it stays countable.

No private-use character in this dataset is an unidentified symbol. The one that used to be — a Wingdings 2 ornament appearing 36 times in the 33rd report, which earlier revisions carried as U+F093 and described as having no Unicode equivalent — is now U+1F668 HOLLOW QUILT SQUARE ORNAMENT. That matters beyond tidiness: the report uses it as a legend marker and defines it in its own text (नोटः 🙨 चिन्हले आंशिक दाबी पुगेका मुद्दाहरूलाई जनाउँछ। — the 🙨 mark indicates cases where the claim was partially upheld), so 34 case-table rows were carrying a documented outcome flag in a codepoint no font could render and no reader could look up.

Page coverage is complete — no page's content is missing.

anchor_fidelity	pages	meaning
tight	8,630	the page's own text is under its own anchor — a citation here is exact
folded	34	text is present but sits under an earlier anchor (see the anchor caveat)
blank	62	the PDF's text layer is genuinely empty — cover, back cover, separator
gap	0	content absent from the dataset

Per-report audit. Transcripts: 35 clean, 0 suspect, 0 garbled on 11 independent axes. Chart data: 7 trustworthy, 6 needs_work.

Charts. 477 charts, 6,116 data points. 277 of 477 charts came through a second independent read unchanged; the other 200 carry a verify_note describing what the second read corrected. 478 of 6,116 values (7.8%) were not printed on the mark and were read off the axis — flagged value_estimated. Chart families: pie 165, bar 165, line 122, stacked_bar 12, venn 5, donut 5, radar 1, other 1, bubble 1.

Tables. 1,735 tables, 86,496 cells parse as numbers.

Known limitations — read these before quoting a number

0. The text was produced two different ways, and coverage is not uniform across the series. 21 reports have a usable PDF text layer and were converted by likhit; 14 reports — the earliest ones — have no text layer at all (zero extractable characters), so every page of them was rendered at 300 DPI and transcribed by a per-page vision read. Filter on documents.text_source (likhit / vision_ocr) whenever provenance matters. Three consequences, all measurable in the data:

   • 2 pages (0.2% of the 1,303 vision-read pages) were NOT transcribed, down from 16: a re-read pass on 2026-08-16 recovered 14 of them, and 10 of those had never been unreadable at all — the page's whole printed content is the annex heading अनुसूचीहरू, the reader returned exactly that, and a length heuristic discarded it as commentary. Two more were reported illegible only because a recorded rotation rendered them upside down. What remains is one blank leaf carrying nothing but a rubber stamp, and one sheet whose 1-bit scan threshold destroyed the Devanagari headstroke (4.3% of its ink in horizontal runs ≥8px, against 20–23% on pages of the same report that transcribed cleanly). They carry anchor_fidelity = 'ocr_unread' with an <!-- ocr:reason --> naming the failure, and are distinguished from genuinely blank pages by measuring inked pixels on the render rather than by trusting the model. This is the dataset's real coverage hole; count it, do not assume the page was empty.

   • Chart data exists only for the 25 reports that had a chart pass. The vision-read reports have no figures rows at all, so a chart in them is absent rather than unreadable. documents.n_figures = 0 is the tell.

   • How well each vision-read page is verified is a column, not a caveat. 1,288 of the 1,288 pages carrying a vision transcription were read more than once, independently, and those reads were compared cell by cell; the other 0 were not, and the columns say which is which rather than leaving a reader to assume. pages.ocr_reads says how many reads that page has, pages.ocr_numeric_cells / pages.ocr_numeric_agreed how many of its numeric cells the reads compared and agreed on, and v_table_numbers carries both onto every numeric cell so numbers can be filtered by evidence in one query:

     -- numeric facts from pages where at least two independent reads agreed
     SELECT * FROM v_table_numbers
     WHERE ocr_source = 'vision' AND ocr_reads >= 2
       AND ocr_numeric_agreed = ocr_numeric_cells;
     

     Measured on this build: 16,293 of 16,634 compared numeric cells agree (97.95%). 100.0% of the vision arm's 12,057 numeric cells sit on pages with two or more reads; the remaining 0.0% (0 cells, 0 pages) have a single read and no agreement figure at all — those are the least verified numbers in the dataset. A further 0 pages carry a vision transcription from an earlier pass with no recorded read count (ocr_reads = 0); they hold no table cells, so they are absent from the figures above rather than silently counted as verified.

   • 🛑 Sideways annex sheets were re-read upright, and that was a real defect. The statistical annexes are landscape sheets the scanner fed sideways into a portrait frame. /Rotate is set uniformly per document so it does not straighten them and the page rectangle stays portrait — nothing in the geometry marks them. The first pass therefore handed the model pages whose text ran vertically, and got structurally broken tables back: one 26-column annex returned a duplicated header carrying two different column sets. 71 pages were found on the rendered pixels, re-read at their correct rotation, and now carry pages.ocr_rotate (90 or 270). Do not compare a ocr_rotate > 0 page against an older release of this dataset and expect the same text.

   • 208 digit corrections were applied, each forced by two independent witnesses — and 155 rows are still reported as not adding up. These annexes print their own totals in both directions, which makes a table a two-dimensional parity check. Where a row failed its own printed जम्मा, a correction was written only when a specific cell's required value was independently confirmed either by another read of the same page or by the column's own printed total; two candidate cells, or none, meant no write. Two of the corrections were checked by hand against 400 DPI crops of the paper and both were right (५०→८० and १३→९३). What could not be attributed is counted, not hidden: pages.ocr_arith_unresolved marks 155 rows over 30 pages whose arithmetic does not close. Some of those are the source's own arithmetic errors — these are hand-set 1990s tables — and at least one is a digit two independent reads both misread the same way.

   • 651 token occurrences over 367 pages were repaired, with the same two-witness rule. Slightly over half are consonant swaps; the rest change a vowel mark or a visarga. pages.ocr_token_corrections is the per-page count, and the run record carries the full breakdown by kind — stated qualitatively here on purpose, because every other figure in this card is derived from the shipped data and a hardcoded ratio would be the one thing in it that could go stale. A vision read of degraded Devanagari swaps visually similar consonants, which no quality axis can see: भ is a valid character, so there is no U+FFFD and no drop in the Devanagari ratio. Measured against the 21 non-OCR reports as a control, the stem बुझ was written बुभ in 71.4% of its occurrences against 0.1% in the control. A token was only rewritten where two independent reads disagreed on it, the variants differ by exactly one character, one is well attested in the control corpus and the other appears there zero times — so घिउ/बिउ (ghee/seed, both real words) is untouched, no token containing a digit is ever eligible, and a token that two pages resolved differently was dropped everywhere. ocr_confusions.json in the corpus repository still lists the candidates that were deliberately not applied. Where both reads agree on a wrong spelling nothing changes, which is the honest limit of the method: अष्टाचार for भ्रष्टाचार is two edits away and stays. Search for both spellings.

   • Width is still the best single predictor of trouble. 32.5% of the vision arm's numeric cells sit in tables wider than 12 columns (20 tables, 3,532 cells) — the sideways annex class above. Join tables.n_cols > 12 to documents.text_source = 'vision_ocr' to isolate or exclude them.

1. A page anchor is a lower bound, never an upper bound. The converter emits a multi-page ruled table as one block under the anchor of the page where the table started, so text from a later page can sit under an earlier anchor. 8,630 of 10,015 pages are tight and 34 are folded; on this build the displacement reaches at least 5 pages (a lower bound — the instrument finds the nearest non-empty preceding anchor, not necessarily the true host block; the upstream corpus measured 24 pages on an earlier conversion). Filter anchor_fidelity = 'tight' when a page citation must be exact, and verify against the PDF before publishing one.

2. Chart coverage is a sample, not an enumeration. Chart pages were found by a detector (curve density, coloured fills, raster coverage), then read. Every page it proposed was read, so nothing was proposed and skipped — but the audit of record bounds the genuinely missed data charts at between 2 and 25, and it is not zero: चित्र ३.१ in the 33rd and चित्र २.३ in the 32nd are confirmed cited-in-prose but untranscribed.

3. Numerals mix Devanagari and ASCII digits — in the CIAA's own typesetting, not in the extraction. 40,५५,५९,२६2, २०७5, २४,085 are verbatim source. 3,231 table cells are affected. A naive ०→0 conversion will mis-parse up to 4.06% of numerals in the worst report. table_cells.value_num already handles both scripts and is NULL wherever the token is not a clean scalar (a fiscal year like २०७९।८०, a section number like 1.३.); text is always verbatim. Do not "fix" the mixed digits — that would falsify the source.

4. 6 reports' chart data is flagged needs_work (2071/72, 2073/74, 2076/77, 2077/78, 2078/79, 2080/81) for internal inconsistencies the reports themselves contain or that the read may have introduced: stacked-bar totals that do not cross-foot against their parts, pie shares that do not close to 100, and the same series reported differently in two reports' five-year trend charts. quality.figure_findings_json names each one. These are not silently corrected.

5. List markers are frequently separated from their item text. In the two-column layouts the converter's fragment ordering strands a list marker from the text it introduces. Derived from the shipped text: 2,130 line-initial - text items against 972 bare - lines carrying no text at all — 31.3% of the 3,102 line-initial markers are orphans. Markers stranded mid-line are not counted here because they are indistinguishable from an ordinary prose hyphen in the output; diffing against the pre-promotion conversion puts total orphaning at 1,066 of 2,133 promoted markers (50.0%), which is the figure to use for the scale of the defect. It loses no characters, so every character-fidelity axis scores it clean. Prose and tables are unaffected; nested-list structure is not reliable.

6. Table headers are not identified. The source tables use multi-row spanned headers with no markdown separator row, so header_json is empty for all but 396 of 1,735 tables and every row lands in the body. intro_text gives the introducing sentence from the prose where there is one; none of the 1,735 tables carries a formal तालिका N caption.

7. The 36th report (FY 2082/83) is not here because the CIAA has not published it. Reports appear 12–18 months after the fiscal year ends.

Provenance and reproduction

• Source: https://ciaa.gov.np/publications — each report's source_url is in documents. PDFs are byte-identical to what was downloaded; pdf_sha256 pins them.
• Text, 21 reports: likhit at revision e40093a — a Nepali PDF→Markdown converter built for legacy pre-Unicode Devanagari font repair. documents.likhit_rev records the revision per report, so any transcript can be traced to the converter build it was produced by — but see Repairs after conversion below before treating that as a reproduction recipe.
• Text, 14 reports: a per-page vision read of 300-DPI renders, because those PDFs carry no text layer for any converter to read. documents.likhit_rev is deliberately empty for them — naming a converter revision on bytes it never produced would be a false provenance stamp. Each page instead carries ocr_src_image and the SHA-256 of the exact image that was read, so a page can be re-read and compared.
• Chart data: two independent vision reads per page with per-value reconciliation, against 300-DPI renders shipped in figure_images/.
• This dataset: assembled directly from the PDFs and transcripts shipped here. MANIFEST.json carries a SHA-256 for all 10,619 files, and documents carries one for every source PDF and transcript, so the whole thing is checkable end to end.
• Repairs after conversion, and this is the part to read before trusting a figure. The transcripts are not raw converter output. Seven recorded passes have edited them in place since conversion: a re-read of every sideways annex page upright; a closure pass on rows whose printed totals did not add up; two font-decode repairs (a dropped repha, and two reports' cmap mis-mappings); virama and Private-Use-Area residue fixes; and a blind cross-instrument second read of the pages that had only been read once. So likhit_rev names the revision a transcript was converted at, and re-running that revision will not reproduce the shipped bytes — the repairs are recorded per page in the ocr: markers and per pass in the project's run records, not in the converter. pages.ocr_reads says how many independent readers saw a vision page, ocr_digit_corrections and ocr_token_corrections how many cells each second read changed, and ocr_arith_unresolved how many a second reader could not settle. Those last ones are shipped rather than hidden: a page with unresolved cells is a page to check against the PDF.

Five reports also have an older transcript produced by a general-purpose converter. Those are deliberately not included: all five score garbled, and a text search in them returns zero hits for words that are plainly present in the report — including, in one case, the passage giving the CIAA's own explanation of its falling bribery conviction rate. Only the repaired transcripts ship here.

Licence and rights

The transcripts, table extractions, chart data and tooling are released under CC BY-NC 4.0.

The source PDFs are government documents — publications of the CIAA, a constitutional body of the Government of Nepal, issued for free public download by the Commission itself. They are republished here unmodified, as public record, so that any extracted value can be checked against its original; documents carries a SHA-256 for each one. Note that 8 of the 35 reports print a rights colophon — the word सर्वाधिकार (all rights) — and 4 of those spell it out in full as सर्वाधिकारः प्रकाशकमा निहित, all rights vested in the publisher. Both counts are given because they are different measurements and an earlier version of this card quoted the full phrase alongside the count for the shorter one. Recorded for completeness; if you have a rights concern, open a discussion on this dataset.

Report content names individuals — officials prosecuted, convicted or acquitted. It is reproduced as published by the CIAA, with no additions. Treat it as the primary legal record it is: an entry does not by itself establish guilt, and outcomes recorded in one year's report may have been overturned on appeal later.

Citation

@misc{ciaa_annual_reports,
  title        = {CIAA Annual Reports: Nepali transcripts, ruled tables and chart data},
  year         = {2026},
  howpublished = {Hugging Face dataset},
  url          = {https://huggingface.co/datasets/damo-da/ciaa-annual-reports},
  note         = {Reports 1--35, BS 2047/48--2081/82; source PDFs published by the
                  Commission for the Investigation of Abuse of Authority, Nepal.
                  Text extracted with likhit e40093a.}
}

Schema

pages

column	type
page_uid	text
doc_id	text
fiscal_year_bs	text
report_ordinal	integer
page_no	integer
text	text
anchor_fidelity	text
coverage_state	text
pdf_text_layer_chars	integer
landscape	integer
n_table_blocks	integer
chars	integer
devanagari_chars	integer
latin_chars	integer
dev_digits	integer
ascii_digits	integer
replacement_chars	integer
control_chars	integer
pua_chars	integer
ocr_source	text
ocr_src_image	text
ocr_date	text
ocr_note	text
ocr_reason	text
ocr_read2	text
ocr_read3	text
ocr_rotate	integer
ocr_reads	integer
ocr_numeric_cells	integer
ocr_numeric_agreed	integer
ocr_digit_corrections	integer
ocr_token_corrections	integer
ocr_arith_unresolved	integer

tables and table_cells

column	type
table_uid	text
doc_id	text
fiscal_year_bs	text
page_no	integer
table_index_on_page	integer
intro_text	text
intro_source	text
intro_is_formal_caption	integer
in_code_fence	integer
header_json	text
n_rows	integer
n_cols	integer
n_cells	integer
n_nonempty_cells	integer
n_numeric_cells	integer
n_mixed_digit_cells	integer
has_separator_row	integer
anchor_fidelity	text
markdown	text

column	type
table_uid	text
doc_id	text
page_no	integer
row_idx	integer
col_idx	integer
text	text
value_num	real
is_numeric	integer
is_empty	integer
beyond_row_width	integer
mixed_digit_scripts	integer

figures and figure_data

column	type
figure_uid	text
doc_id	text
fiscal_year_bs	text
page_no	integer
figure_index_on_page	integer
figure_label	text
title_ne	text
chart_type	text
chart_family	text
x_axis	text
y_axis	text
unit	text
notes	text
verify_note	text
verified_clean	integer
n_data_points	integer
n_estimated_values	integer
page_image	text

column	type
figure_uid	text
doc_id	text
page_no	integer
point_index	integer
label	text
series	text
value	real
value_estimated	integer

documents and quality

column	type
doc_id	text
report_ordinal	integer
fiscal_year_bs	text
ad_year_start	integer
ad_year_end	integer
title_ne	text
source_url	text
pdf_file	text
pdf_sha256	text
pdf_bytes	integer
pdf_page_count	integer
pdf_producer	text
markdown_file	text
markdown_sha256	text
markdown_bytes	integer
n_pages	integer
pages_anchored	integer
pages_blank	integer
pages_folded	integer
pages_gap	integer
pages_ocr_vision	integer
pages_ocr_unread	integer
pages_with_text	integer
n_tables	integer
n_figures	integer
n_figure_data_points	integer
n_estimated_figure_values	integer
figures_quality_verdict	text
transcript_quality_verdict	text
text_source	text
likhit_rev	text
text_chars	integer
text_devanagari_chars	integer
text_latin_chars	integer
text_dev_digits	integer
text_ascii_digits	integer
text_replacement_chars	integer
text_control_chars	integer
text_pua_chars	integer

column	type
doc_id	text
fiscal_year_bs	text
transcript_verdict	text
transcript_failing_axes	text
transcript_dev_ratio	real
n_axes_checked	integer
axes_json	text
figures_quality_verdict	text
n_figure_findings	integer
figure_findings_json	text

Per-report coverage

Report	FY (BS)	FY (AD)	pages	chars	tables	charts	chart points	transcript	chart data
1st	2047/48	1990/91	11	19,280	7	0	0	clean	``
2nd	2048/49	1991/92	26	52,308	9	0	0	clean	``
3rd	2049/50	1992/93	36	95,306	11	0	0	clean	``
4th	2050/51	1993/94	91	153,077	11	0	0	clean	``
5th	2051/52	1994/95	59	98,187	10	0	0	clean	``
6th	2052/53	1995/96	66	200,140	20	0	0	clean	``
7th	2053/54	1996/97	172	594,974	34	0	0	clean	``
8th	2054/55	1997/98	88	328,355	28	0	0	clean	``
9th	2055/56	1998/99	95	283,889	25	4	28	clean	``
10th	2056/57	1999/00	111	379,851	26	2	10	clean	``
11th	2057/58	2000/01	106	355,200	23	6	154	clean	``
12th	2058/59	2001/02	129	446,744	31	0	0	clean	``
13th	2059/60	2002/03	107	305,159	44	2	21	clean	``
14th	2060/61	2003/04	206	494,637	59	6	43	clean	``
15th	2061/62	2004/05	336	811,656	35	8	88	clean	``
16th	2062/63	2005/06	349	915,175	32	1	32	clean	``
17th	2063/64	2006/07	321	775,725	33	12	149	clean	``
18th	2064/65	2007/08	317	848,125	27	12	148	clean	``
19th	2065/66	2008/09	444	1,129,717	41	12	161	clean	``
20th	2066/67	2009/10	471	1,229,159	17	0	0	clean	``
21st	2067/68	2010/11	406	1,015,392	35	12	174	clean	``
22nd	2068/69	2011/12	407	902,525	40	12	182	clean	``
23rd	2069/70	2012/13	542	1,533,948	44	12	180	clean	trustworthy
24th	2070/71	2013/14	865	2,629,245	36	6	76	clean	trustworthy
25th	2071/72	2014/15	359	1,000,854	35	18	229	clean	needs_work
26th	2072/73	2015/16	363	1,008,174	39	13	135	clean	trustworthy
27th	2073/74	2016/17	346	978,175	42	14	151	clean	needs_work
28th	2074/75	2017/18	255	834,158	41	19	187	clean	trustworthy
29th	2075/76	2018/19	328	1,045,489	60	31	481	clean	trustworthy
30th	2076/77	2019/20	364	1,144,386	72	40	480	clean	needs_work
31st	2077/78	2020/21	296	924,507	86	38	467	clean	needs_work
32nd	2078/79	2021/22	461	1,029,401	151	44	580	clean	needs_work
33rd	2079/80	2022/23	492	1,134,283	250	49	609	clean	trustworthy
34th	2080/81	2023/24	544	1,220,883	136	58	771	clean	needs_work
35th	2081/82	2024/25	446	1,131,589	145	46	580	clean	trustworthy