W4ashabii/Gov_QA_OPUS

Nepal Gov QA: Opus smoke run

Synthetic Nepali search queries for chunks of Nepali government web pages and PDFs, each paired with the chunk that answers it. Built for training retrieval and embedding models.

This is a 300-chunk smoke run. The queries were written by Claude Opus 5.5, and the chunks were pre-filtered for text quality and personal data. It replaces the rejected W4ashabii/Gov_QA_test run, which used gemma3:4b and unfiltered text.

⚠️ Not fully verified. No person has reviewed these pairs, and an automated check showed that some queries ask for details the passage doesn't contain. Read Verification below before training on this set.

Source corpus: voidash/previllage-nepal-gov-corpus. Most of its chunks have no title (92.5% of chunks, 51.8% of documents), which is why generated queries are useful here.

Configs

pairs (default): 283 query–passage pairs over 159 chunks.

field	description
query	Nepali query written by Opus (median 56 characters)
passage	The chunk text, exactly as the model saw it (site breadcrumbs removed, whitespace collapsed)
chunk_id, doc_id, source_id	IDs in the source corpus
document_url	URL of the original page or PDF
has_title	Whether the chunk has a title in the source corpus (105 of 283 pairs)
bge_m3_rank	Rank of this passage when the query is searched with BAAI/bge-m3 against the 300 chunks plus 19,966 random corpus chunks. Recorded for information; nothing was filtered on it.
verifier_checked	Whether the pair was in the automated support check (89 pairs; see Verification)
model	claude-opus-5-5

all_queries: all 300 chunks with their status.

status	chunks	meaning
ok	159	Has queries in pairs
unusable	102	Opus found nothing worth searching for, e.g. menus, bare tables or fragments
dropped_garbled_text	39	Queries were written, but the passage turned out to use a wrong legacy-font mapping, e.g. "कायाुन्द्वयन गनु"

from datasets import load_dataset
pairs = load_dataset("W4ashabii/Gov_QA_OPUS", "pairs", split="train")

How it was made

1. Eligible chunks. Keep chunks of current documents, drop exact-duplicate texts and drop chunks under 20% Devanagari. That leaves 185,583 chunks.

2. Quality and personal-data filter. Drop chunks with any of the following:

   • unconverted Preeti font encoding (e.g. @)*!÷)*@)
   • phone numbers or email addresses
   • private-use glyphs
   • split vowel signs from bad PDF extraction
   • stray Latin text
   • mostly digits or very few words

   Only 24% of eligible chunks (45,113) pass.

3. Sample. 300 clean chunks, one per source (300 sources), none from the sources used in the rejected run. 58% have no title.

4. Generation. Claude Opus 5.5 read the full chunk text and wrote 1–2 specific Nepali queries, or marked the chunk unusable. 364 queries came back; the cleaner then dropped 2 (length, script or verbatim-copy checks).

5. Post-filter. A second font-garbling check dropped 39 chunks (75 pairs) that step 2 had missed. This check is now part of the step-2 filter for future runs.

Retrieval sanity check (bge-m3, ~20k-chunk pool): the query's own passage ranks first for 52% of queries, top-3 for 70% and top-10 for 80%. The median rank is 1.

Verification

• Spot checks found 4 unsupported queries, all removed. One named the right municipality but asked about a programme from a different chunk ("बगर खेतीका लागि अनुदान"). Three asked for details the passage lacks: a named authority, a site-specific budget, and a fixed fee.
• The automated check was too lenient. 92 suspicious pairs (bge-m3 rank above 5, or few words shared with the passage) went to an LLM verifier. It passed all 92, including the 3 just mentioned, so it cannot be trusted to catch everything.
• The flagging missed at least one invented query. The "बगर खेती" query ranked 3rd and was never flagged. Pairs with verifier_checked = false are therefore not known to be clean.
• Navigational pairs. Some pairs are searches for a named document whose passage only lists that document's title (for the right office). For a "find this document" search that is a correct result, so these were kept.

Bottom line: most pairs look correct, but some unsupported queries remain. Review the pairs or check them more strictly before use, especially for evaluation.

Known limitations

• Coverage. Only 300 chunks, with heavy filtering: 53% of sampled chunks produced pairs.
• Undetected garbling. Some passage text may still be garbled in ways the filters don't detect.
• Coverage bias. Strict filtering favours chunks from sites with clean Unicode text, so older PDFs typed in legacy fonts are under-represented.
• Personal data. Chunks with phone numbers or emails were removed, but names of officials can still appear in passages.

Licence and attribution

Released under CC BY 4.0, following the source corpus. That licence covers the compilation and the generated queries. The passage text comes from Nepali government websites and keeps whatever rights and terms the original pages carry. Check the source at document_url before reusing passage text, and credit the original source.