SayedShaun/bangla-retrieval-collection

Bangla Retrieval Collection

A unified collection of Bangla (Bengali) query–passage training data for fine-tuning embedding and reranker models, gathered from 6 public datasets into a single schema.

This repository is a collection, not a new dataset. All credit for the underlying data belongs to the original authors (see Sources & credit). Each subset keeps its original license. What this repository adds is Bangla-only filtering, a common format, removal of evaluation queries, and deduplication.

Quick start

from datasets import load_dataset

ds = load_dataset("SayedShaun/bangla-retrieval-collection", split="train")                   # all subsets
ds = load_dataset("SayedShaun/bangla-retrieval-collection", "commercial-ok", split="train")  # excludes NC subsets
ds = load_dataset("SayedShaun/bangla-retrieval-collection", "miracl", split="train")         # one subset

Schema

column	type	description
query	string	Bangla query
positive	string	relevant passage (title text where the source has titles)
negative	string or null	hard negative passage; null for pair-only subsets
source	string	subset name

Subsets

subset	rows	unique queries	has negatives	type	license
miracl	380,760	1,624	yes (~234 per query)	native Bangla, human-labelled	Apache-2.0 (passages: Wikipedia, CC-BY-SA)
mr-tydi	168,405	1,693	yes (~100 per query)	native Bangla, human-labelled	Apache-2.0 (passages: Wikipedia, CC-BY-SA)
swim-ir	104,391	101,639	no	native passages, LLM-generated queries	CC-BY-SA-4.0
squad-bn	69,592	69,542	no	machine-translated SQuAD	CC-BY-NC-SA-4.0
banglarqa	10,119	10,111	no	native Bangla reading comprehension	CC-BY-NC-SA-4.0
miracl-nlpai	2,859	2,843	yes (1 per query)	native Bangla, human-labelled	Apache-2.0 (passages: Wikipedia, CC-BY-SA)
total	736,126

miracl and mr-tydi repeat each (query, positive) pair with many different negatives. To train with a balanced mix, cap the number of rows per query. squad-bn and banglarqa reuse each passage for several questions, so use a no-duplicates batch sampler when training with in-batch negatives.

Licensing

This is a compilation. Every subset is distributed under the license of its original source (table above). It is not relicensed as a whole.

• squad-bn and banglarqa are non-commercial (CC-BY-NC-SA-4.0). For commercial use, load the commercial-ok config, which contains only miracl, mr-tydi, miracl-nlpai and swim-ir.
• Share-alike subsets (CC-BY-SA, CC-BY-NC-SA) require that redistributed adaptations keep the same license.
• The passages in the MIRACL and Mr.TyDi subsets come from Bangla Wikipedia (CC-BY-SA).

If you are an author of a source dataset and want your data removed or its attribution changed, please open a discussion on this repository.

Processing

1. Conversion: each source is converted to (query, positive, negative). QA datasets use question → context. For BanglaRQA, only answerable questions are kept.
2. Normalization: Unicode NFC, and all whitespace (including newlines) collapsed to single spaces.
3. Bangla-only filter: at least 60% of letters in each text must be in the Bengali Unicode block. Assamese text is removed: a text is treated as Assamese when the Assamese-only letters ৰ/ৱ make up more than 0.5% of its Bengali-script characters.
4. Evaluation decontamination: queries from MIRACL bn dev (mteb/MIRACLReranking) and Mr.TyDi bn dev/test are removed. The SQuAD-bn test and BanglaRQA Test splits are not included, so they stay usable for evaluation.
5. Deduplication: exact duplicates are removed within each subset, and so are rows where positive == negative.

Recommended evaluation

mteb/MIRACLReranking (bn), mteb/MIRACLRetrieval (bn), mteb/IndicQARetrieval (bn), the Mr.TyDi bn test set, and the SQuAD-bn / BanglaRQA test splits.

Compatible data not included

These Bangla MS MARCO translations work with the same pipeline, but they are not redistributed here because they state no license. Load them from their original repositories:

• kazalbrur/MSMARCO-BN: about 5.0M triplets after processing
• samikhan121/indicmarco-triples: about 25.4M triplets after processing

Sources & credit

subset	original dataset	paper
miracl	miracl/miracl, via the triplet format of sentence-transformers/miracl	Zhang et al., 2023 (TACL)
mr-tydi	castorini/mr-tydi, via sentence-transformers/mr-tydi	Zhang et al., 2021
miracl-nlpai	nlpai-lab/miracl-multilingual-triplets	derived from MIRACL
swim-ir	nthakur/swim-ir-monolingual	Thakur et al., 2023
squad-bn	csebuetnlp/squad_bn	Bhattacharjee et al., 2021 (BanglaBERT)
banglarqa	sartajekram/BanglaRQA	Ekram et al., 2022

Thanks to all original authors, and to the Sentence Transformers team for the MIRACL and Mr.TyDi triplet versions.

Citations

If you use this collection, please cite the original datasets for the subsets you use:

@article{zhang2023miracl,
  title   = {{MIRACL}: A Multilingual Retrieval Dataset Covering 18 Diverse Languages},
  author  = {Zhang, Xinyu and Thakur, Nandan and Ogundepo, Odunayo and Kamalloo, Ehsan and Alfonso-Hermelo, David and Li, Xiaoguang and Liu, Qun and Rezagholizadeh, Mehdi and Lin, Jimmy},
  journal = {Transactions of the Association for Computational Linguistics},
  volume  = {11},
  pages   = {1114--1131},
  year    = {2023}
}

@article{mrtydi,
  title   = {{Mr. TyDi}: A Multi-lingual Benchmark for Dense Retrieval},
  author  = {Xinyu Zhang and Xueguang Ma and Peng Shi and Jimmy Lin},
  year    = {2021},
  journal = {arXiv:2108.08787}
}

@article{thakur:2023,
  author  = {Nandan Thakur and Jianmo Ni and Gustavo Hern{\'{a}}ndez {\'{A}}brego and John Wieting and Jimmy Lin and Daniel Cer},
  title   = {Leveraging LLMs for Synthesizing Training Data Across Many Languages in Multilingual Dense Retrieval},
  journal = {CoRR},
  volume  = {abs/2311.05800},
  year    = {2023},
  doi     = {10.48550/ARXIV.2311.05800}
}

@misc{bhattacharjee2021banglabert,
  title         = {BanglaBERT: Combating Embedding Barrier in Multilingual Models for Low-Resource Language Understanding},
  author        = {Abhik Bhattacharjee and Tahmid Hasan and Kazi Samin and Md Saiful Islam and M. Sohel Rahman and Anindya Iqbal and Rifat Shahriyar},
  year          = {2021},
  eprint        = {2101.00204},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL}
}

@inproceedings{ekram-etal-2022-banglarqa,
  title     = {{B}angla{RQA}: A Benchmark Dataset for Under-resourced {B}angla Language Reading Comprehension-based Question Answering with Diverse Question-Answer Types},
  author    = {Ekram, Syed Mohammed Sartaj and Rahman, Adham Arik and Altaf, Md. Sajid and Islam, Mohammed Saidul and Rahman, Mehrab Mustafy and Rahman, Md Mezbaur and Hossain, Md Azam and Kamal, Abu Raihan Mostofa},
  booktitle = {Findings of the Association for Computational Linguistics: EMNLP 2022},
  year      = {2022},
  address   = {Abu Dhabi, United Arab Emirates},
  publisher = {Association for Computational Linguistics},
  url       = {https://aclanthology.org/2022.findings-emnlp.186},
  pages     = {2518--2532}
}