momahadi/bangladesh-legal-qa-dataset

Bangladesh Legal QA Dataset: Bangla-English Law and Fine-Tuning

The Bangladesh Legal QA Dataset is a bilingual Bangla-English dataset for Bangladesh law question answering, legal NLP, LLM fine-tuning, instruction tuning, and retrieval-augmented generation (RAG). It provides 2,165 context-grounded legal QA records, direct-answer and IRAC chat-format training data, and structured statutory text from six Bangladesh Acts and three schedules.

This is the 2,165-record paper-aligned release used for the associated small-language-model legal QA study. It does not include the separate, later 3,519-record dataset family.

Associated paper: *Do Small Models Use the Law You Give Them? Measuring Context Use on a Bilingual Bangladesh Legal Benchmark*.

Separate Bangladesh Bar Council benchmark

The 400-question 2022-2023 Bangladesh Bar Council evaluation benchmark is published separately:

[Bangladesh Bar Council Exam QA Dataset](https://huggingface.co/datasets/momahadi/bangladesh-bar-council-exam-dataset)

The separation keeps the open license for this research dataset clear while allowing the exam repository to carry its own source-rights notice. Files named bar_exam_style_* in this repository are researcher-created QA training subsets, not the raw Bar Council examination files.

What the dataset contains

QA record breakdown

Repository structure

.
|-- sft/
|   |-- finetune_dataset_2165.json
|   |-- finetune_answer_only_2165.jsonl
|   `-- finetune_irac_answer_2165.jsonl
|-- qa-splits/
|-- law-corpus/
|   |-- english/
|   |-- bangla/
|   `-- schedules/
|-- CONTRIBUTORS.md
|-- CITATION.bib
|-- LICENSE.md
`-- VALIDATION.md

Files

SFT data

QA splits

Loading the data

from datasets import load_dataset

repo = "momahadi/bangladesh-legal-qa-dataset"

audit_records = load_dataset(repo, "audit", split="train")
answer_only = load_dataset(repo, "direct_answer", split="train")
irac = load_dataset(repo, "irac", split="train")

Load the separate examination benchmark with:

bar_repo = "momahadi/bangladesh-bar-council-exam-dataset"
bar_2022_bn = load_dataset(bar_repo, "bar_council", split="test_2022_bn")

Dataset fields

The audit-friendly JSON contains fields for dataset ID, language, question type, Act and section metadata, candidate legal sections, question, answer, IRAC reasoning, quality metadata, and source-file provenance. Some fields are present only for particular question types.

Each JSONL training record has one messages array using the standard system/user/assistant chat structure.

Intended uses

• —Bangladesh legal NLP and bilingual Bangla-English legal QA;
• —supervised fine-tuning and instruction-tuning experiments;
• —IRAC-formatted legal-answer generation;
• —retrieval-augmented and context-grounded legal QA;
• —controlled evaluation of small language models;
• —low-resource-language and cross-lingual legal AI research.

Limitations

• —This is a research dataset, not legal advice.
• —Answers and legal references may contain errors and should be independently verified before high-stakes use.
• —The records cover a limited set of statutes, schedules, years, languages, and question types.
• —The 2,165-record release must not be merged silently with the later 3,519-record dataset family.

Provenance

The bundle was assembled from the paper-aligned research artifacts:

• —structured legal context from six Bangladesh Acts and three schedules;
• —final QA splits for single-hop, advanced-selection, and bar-exam-style questions;
• —final direct-answer and IRAC SFT variants.

Model-result files and score tables are not included.

Contributors and research supervision

The repository is maintained under a personal Hugging Face account, but credit is attached to the dataset itself. The ordered release list is:

1. 1.Moniruzzaman Mahadi - dataset author and corresponding maintainer
2. 2.Abrar Mohammed Tanzim Alam - dataset contributor
3. 3.Sayma Siddika Monalisa - dataset contributor
4. 4.Mir Mohammad Asif Abdullah - dataset contributor
5. 5.Mahina Rahman Deya - dataset contributor
6. 6.Swakkhar Shatabda - research supervisor
7. 7.Md Adnan Arefeen - research supervisor

Correspondence: momahadi9664@gmail.com

See `CONTRIBUTORS.md` for contributor details and `CITATION.bib` for the paper citation. More detailed CRediT roles can be added after confirmation by the team.

Citation

Cite the paper. There is no separate citation for this dataset.

@misc{mahadi2026smallmodelsuselaw,
  title={Do Small Models Use the Law You Give Them? Measuring Context Use on a Bilingual Bangladesh Legal Benchmark},
  author={Moniruzzaman Mahadi and Abrar Mohammed Tanzim Alam and Sayma Siddika Monalisa and Mir Mohammad Asif Abdullah and Swakkhar Shatabda and Md Adnan Arefeen},
  year={2026},
  eprint={2608.30327},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2608.30327},
}

Attribution under CC BY 4.0 is satisfied by citing the paper. Dataset contributors are credited in CONTRIBUTORS.md.

License and source rights

The release team's original dataset contributions are licensed under the Creative Commons Attribution 4.0 International license (CC BY 4.0). This includes the original questions, answers, annotations, metadata, selection, arrangement, and researcher-produced translations to the extent the release team holds rights in them.

Official statutory text is source material and was not authored by the release team. The team does not claim ownership of that text or purport to relicense rights it does not hold. See `LICENSE.md` for the exact scope and attribution requirements.