MYOTHANTZIN/eps-burmese-qa

EPS Burmese Legal QA

Burmese-language question–answer pairs about Korean labour and immigration law, grounded in the statutes themselves, for Myanmar workers on E-9/EPS visas in South Korea.

Korean employment law governs the daily life of hundreds of thousands of migrant workers. The statutes exist in Korean and in official English translation. Almost none of it exists in Burmese. This dataset was built to change that, and to make it possible to measure whether a model answers such questions correctly.

Burmese evaluation data is scarce. The test split is offered as a small benchmark for grounded QA in a low-resource language, where answers must cite a source and decline when the law does not cover the question.

Splits

split	rows	grounded	refusal
train	522	473	49
test	50	45	5

The two splits are disjoint and their union is the complete 572-row set. They are published separately rather than as a third combined split to avoid shipping every row twice.

from datasets import load_dataset
ds = load_dataset("MYOTHANTZIN/eps-burmese-qa")
ds["train"][0]

Fields

field	description
id	unique row id (p001, d702, s012, …)
kind	grounded — the context answers the question · refusal — it does not
system	system prompt, byte-identical across all rows
context	the legal passage the answer is grounded in, prefixed with its source tag
question	the question, in Burmese, phrased as a worker would ask it
answer	the answer, in Burmese, citing its source
orig_id	id in the pre-deduplication version

Why refusals matter

49 of 522 training rows are refusals — questions the provided context genuinely cannot answer, where the correct response is to decline and point the worker at 고용센터 (employment centre) or 출입국관리사무소 (immigration office).

This is the safety-critical behaviour for a legal assistant. A model that answers everything confidently is more dangerous than one that admits ignorance, because a migrant worker acting on an invented rule can lose their visa or their wages.

The refusals use 15 different phrasings across three registers — formal, terse and conversational — written by a native speaker and listed in refusal_variants.txt. An earlier version repeated a single refusal string 48 times, which taught a model to copy that string rather than to recognise not knowing.

Construction

1. Corpus. Korean labour and immigration statutes in official English translation, plus HiKorea procedural pages, chunked at 500 tokens with provenance on every chunk.
2. Retrieval. Candidate passages found with BAAI/bge-m3, then hand-verified — automatic top-k was not trusted.
3. Answers. Written in Burmese against the verified passage, with Korean official terms kept verbatim so a worker can use them at the office.
4. Verification. Every legal citation was grep-checked against the fetched source text. Rows whose cited article could not be found were corrected or dropped.

Some drafts were produced with the help of another model, then verified against sources before inclusion; rows that failed verification were discarded. The method is documented in the source repository.

Benchmark results

results/full50_base_vs_tuned.json contains 100 generations — the base model and a QLoRA fine-tune of it, answering the same 50 test questions with identical prompts and decoding.

	base	fine-tuned
Keeps Korean official terms	5/30	18/30
Cites a source	26/45	34/45
Refuses correctly	2/5	4/5
Article number correct, when cited	13/13	20/21
Stopped naturally	44/50	44/50

Reproduce these numbers with scripts/verify_results.py — it recomputes every figure from the results file and asserts the two conditions are not accidentally the same model.

Adapter: MYOTHANTZIN/eps-burmese-sealion-9b-lora

Limitations

• Small. 572 rows. Enough to teach answer conventions, not enough to teach a domain.
• Short questions. Median question length is 52 characters. Real users write 90–170, and a model trained on this refuses longer questions even when the answer is in the context. If you train on this, add longer conversational phrasings.
• Narrow coverage. E-9/EPS core topics only — workplace change, contracts, wages, insurance, residence. Not tax, housing, healthcare access, schooling or driving.
• One system prompt across every row, so a model trained on it may be sensitive to changes in that prompt's wording.
• Point-in-time. Sources were fetched on 2026-09-01. Korean law changes; minimum wage changes annually.

Provenance

Contexts are excerpts of publicly available Korean government material, each row tagged with its source. Full manifest with URLs, authority and per-source notes in sources.json.

source	authority
law.go.kr	Korea Ministry of Government Legislation
elaw.klri.re.kr	Korea Legislation Research Institute
hikorea.go.kr	Korea Immigration Service
easylaw.go.kr	Ministry of Government Legislation — Easy-to-Find Practical Law

Korean government works are generally published under the Korea Open Government License, which requires attribution; KOGL has several types, so consult each source's own terms for your use case.

All English legal text here is unofficial translation with no legal force. The Korean original governs. The Burmese questions and answers are the author's own work.

Intended use

Research and evaluation: grounded QA, citation behaviour, refusal behaviour, and cross-lingual retrieval in a low-resource language.

Not legal advice, and not fit for advising real workers unmodified. An end-to-end system built on this dataset produced confidently wrong answers roughly 8% of the time, with correct formatting and a plausible citation. Verify with 고용센터 or 출입국관리사무소.

Source

Full methodology, evaluation and the two library bugs found along the way: https://github.com/MYOTHANTZIN-THANT/EPS-Burmese-Assistant

Citation

@misc{eps_burmese_qa_2026,
  title  = {EPS Burmese Legal QA: grounded question answering on Korean labour law in Burmese},
  author = {Myo Thant Zin},
  year   = {2026},
  url    = {https://huggingface.co/datasets/MYOTHANTZIN/eps-burmese-qa}
}