W4ashabii/Gov_QA_test

Nepal Gov QA: smoke run (REJECTED)

⚠️ Rejected. Do not use. This smoke run failed review. Quality checks showed that 153 of the 324 pairs (47%) have at least one problem:

• Unconverted Preeti font encoding (68 pairs), e.g. @)*!÷)*@ in place of a year
• Stray Latin text (45)
• Private-use glyphs (31)
• Broken letter combinations from PDF extraction (29)
• Personal names and phone numbers (13)

Many queries are also vague or off-target, because a small model (gemma3:4b) generated them from noisy text. The bge-m3 round-trip filter did not catch these, since garbled queries still match garbled passages.

Personal data removed: 14 chunks containing phone numbers or email addresses (19 pairs) were deleted from both files. The counts below reflect the cleaned files.

It is superseded by W4ashabii/Gov_QA_OPUS, which pre-filters the source chunks for text quality and personal data and generates queries with Claude Opus 5.5. This repo is kept only as a record of the failed run.

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

This is a 500-chunk smoke run of the generation pipeline, released so its quality can be inspected before a full run. All queries are machine-generated by a small local model and have not been reviewed by people.

Source corpus: voidash/previllage-nepal-gov-corpus. Most of its chunks have no title (92.5% of chunks, 51.8% of documents), so generated queries are the main way to give them a searchable handle.

Configs

pairs (default): 305 filtered query–passage pairs over 231 chunks.

field	description
query	Generated Nepali query
passage	Full text of the source chunk
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
gold_rank	Rank of the source chunk when the query was searched with bge-m3 (1–3)
model	Generator model

all_queries: 486 chunks (500 minus 14 removed for personal data), with every query before the retrieval filter (812 queries).

field	description
status	ok (412 chunks), unusable (59, the model found nothing searchable), prefiltered (15, removed before generation as mostly digits or too few Nepali words)
queries	0–2 queries

The other fields are the same as in pairs, minus passage and gold_rank.

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

How it was made

1. Sampling. From the source corpus, keep chunks of current documents, drop exact-duplicate texts (33,447), and drop chunks under 20% Devanagari (58,875). That leaves 185,583 eligible chunks. The 500 smoke-run chunks come from 500 different sources (298 PDF, 202 HTML); 62.6% have no title.
2. Generation. gemma3:4b (via Ollama, local GPU) gets the first ~1,000 characters of each chunk, cut at a sentence end. The prompt asks for 1–2 specific Nepali queries a citizen would search, or usable=false, and includes one worked example. Output is constrained to JSON.
3. Cleaning. Queries are dropped if they are under 8 or over 150 characters, under 60% Devanagari, contain a Latin word that is not in the chunk, or are copied verbatim from the chunk.
4. Round-trip filter. Each query is embedded with BAAI/bge-m3 and searched against a pool of the 500 source chunks plus 20,000 random eligible chunks. A pair is kept only if its own chunk ranks in the top 3. 324 of 840 queries (39%) passed: 241 ranked first, 63 second and 20 third. After the personal-data removal, 305 pairs remain (225 first, 60 second, 20 third).

Known issues

• Spelling and grammar. A 4B model writes imperfect Nepali, and some queries contain misspellings or awkward phrasing.
• Truncation. Queries come from the first ~1,000 characters of each chunk, so later parts of a chunk are under-covered.
• Near-duplicate pages. Many municipality pages are close to identical. The filter drops general queries that match many of them, so the kept set leans towards specific, named queries.
• Personal data. Chunks with phone numbers or emails were removed, but passages may still name individuals.

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.