QuangDuy/supervised_data-hn-rerank-denoise

HUIT-Embedding — Teacher-denoised hard negatives (Vietnamese)

Hard negatives mined for QuangDuy/supervised_data with the method of RocketQA and RLHN. 17 configs. Generated on 2026-10-25. Candidates come from QuangDuy/supervised_data-hn-candidates.

Method

Every query's positive and top-20 retrieval candidates are scored by the teacher BAAI/bge-reranker-v2-m3 (probability of relevance). Candidates are then split:

• Extra positives (extra_positive_ids): the positive's own text in another format (word 5-gram containment ≥ 0.8 and teacher ≥ 0.9), common in the law corpus.
• Suspected false negatives (suspected_positive_ids): other candidates scoring ≥ 0.9 (RocketQA) or at least the positive. Kept for inspection; use them neither as negatives nor as positives.
• Negatives (negative_ids): every other candidate, i.e. scored below 0.9 and below the positive.

All teacher scores are kept (candidate_ids, candidate_scores) for KL distillation. Rows whose positive the teacher scores < 0.5 are kept and marked positive_disputed; filter them if you want only teacher-confirmed pairs.

• Retriever: intfloat/multilingual-e5-large (paper), exact search over a per-domain corpus (law, NLI, reading comprehension, health, fact-check, tables) built from the positives and provided negatives of the source data, with evaluation overlaps removed.
• Teacher: BAAI/bge-reranker-v2-m3 (paper).
• Scores in negative_scores: teacher probability.

Differences from the paper

• RLHN relabels with an LLM judge; here a reranker flags. A hand check of 30 strongly flagged law candidates found about 70% true false negatives and 30% topical non-answers, so flagged candidates are not turned into positives (only exact-content near-copies are).
• RLHN drops queries with more than 7 flagged candidates; that would remove 59–69% of the law queries, so no query is dropped and the count is reported.
• mmarco_vi uses the hotchpotch scores of 64 BM25 negatives (computed on the unicamp-dl/mmarco text).

Two layouts

• <config>: texts included (query, positive, negatives, suspected_positives). Ready to train.
• <config>-ids: ids and scores only; texts are in documents-<group> (doc_id, text). Much smaller.

Usage

from datasets import load_dataset

ds = load_dataset("QuangDuy/supervised_data-hn-rerank-denoise", "csconda", split="train")

# ids layout: join negatives with the documents table
ids = load_dataset("QuangDuy/supervised_data-hn-rerank-denoise", "csconda-ids", split="train")
docs = load_dataset("QuangDuy/supervised_data-hn-rerank-denoise", "documents-health", split="train")
text = dict(zip(docs["doc_id"], docs["text"]))
row = ids[0]
negatives = [text[d] for d in row["negative_ids"]]

Training: sample 7 of the negatives per step (as in LightOn and NV-Retriever) and keep rows with the same group_id out of one batch. For KL distillation, softmax the teacher scores of [positive] + candidates with a temperature and match the student's distribution over the same texts.

Configs

Config	Source	Queries in	Queries out	Kept	Negatives / query	Queries with flags	Flags / query	Queries with > 7 flags	Extra positives	Disputed positives
csconda	csconda	19	19	100.0%	16.7	68.4%	3.32	4	0	73.7%
isedsc2024	isedsc2024	7,944	7,944	100.0%	19.5	15.1%	0.48	138	10	8.4%
large_vi_legal_queries	large_vi_legal_queries	494,312	494,312	100.0%	8.2	94.4%	10.76	320,189	538,132	2.3%
mmarco_vi	mmarco_vi	480,133	480,133	100.0%	48.0	88.8%	15.58	256,816	19,351	17.7%
multi_wiki_qa	multi_wiki_qa	4,932	4,932	100.0%	19.7	7.1%	0.34	80	30	11.6%
table_markdown_retrieval	table_markdown_retrieval	125,830	125,830	100.0%	15.9	56.2%	4.06	27,602	0	12.8%
vianli	vianli	1,606	1,606	100.0%	15.6	85.6%	4.44	335	0	17.9%
vicov19qa	vicov19qa	2,911	2,911	100.0%	8.8	90.6%	10.99	1,788	563	8.2%
vietnamese_law_hn_mined	vietnamese_law_hn_mined	327,190	327,190	100.0%	9.9	88.9%	8.94	177,521	383,372	2.7%
vifactcheck	vifactcheck	1,595	1,595	100.0%	18.6	49.7%	1.24	34	227	1.1%
vihealthqa	vihealthqa	6,582	6,582	100.0%	13.5	70.0%	6.40	2,292	902	21.1%
vimedaqa	vimedaqa	31,437	31,437	100.0%	17.3	55.2%	2.65	3,641	151	15.6%
vimqa	vimqa	7,158	7,158	100.0%	19.6	24.4%	0.40	19	245	1.0%
vinli	vinli	2,107	2,107	100.0%	17.1	75.3%	2.85	202	5	1.6%
vinli_simcse_v2	vinli_simcse_v2	268,871	268,871	100.0%	8.0	93.6%	11.96	180,624	21,929	9.5%
virhe4qa	virhe4qa	75	75	100.0%	20.0	2.7%	0.03	0	0	13.3%
viwikifc	viwikifc	1,802	1,802	100.0%	19.8	11.5%	0.16	1	2	2.1%

Sources without corpus hard negatives (context-dependent questions: top candidates usually answer them too): vire4mrc, vlogqa. Use them with in-batch negatives from the base dataset.

Sources and licenses

Source	Origin	License	Label origin	Translated
csconda	ura-hcmut/Vietnamese-Customer-Support-QA	cc-by-nc-4.0	human_paper	False
isedsc2024	MiuN2k3/isedsc.train.dataset	unknown	competition	False
large_vi_legal_queries	phamson02/large-vi-legal-queries	cc-by-4.0	synthetic	False
mmarco_vi	unicamp-dl/mmarco, GreenNode/msmarco-vn	apache-2.0 (repo); passages from GreenNode/msmarco-vn cc-by-sa-4.0; MS MARCO terms (non-commercial)	translated	True
multi_wiki_qa	alexandrainst/multi-wiki-qa (config vi)	cc-by-nc-sa-4.0	synthetic	False
table_markdown_retrieval	GreenNode/GreenNode-Table-Markdown-Retrieval-VN	mit	synthetic	False
vianli	uitnlp/ViANLI	mit	human_paper	False
vicov19qa	https://raw.githubusercontent.com/triet2397/UIT-ViCoV19QA/main/dataset/4_ans/UIT-ViCoV19QA_train.csv	unknown (README: research purposes only)	human_paper	False
vietnamese_law_hn_mined	GreenNode/Vietnamese-Law-hn-mined-v2	unknown	competition	False
vifactcheck	tranthaihoa/vifactcheck	mit	human_paper	False
vihealthqa	tarudesu/ViHealthQA	unknown	human_paper	False
vimedaqa	tmnam20/ViMedAQA	unknown	human_paper	False
vimqa	nguyenlab/vimqa	cc-by-nc-sa-4.0	human_paper	False
vinli	presencesw/vinli_4_label	unknown	human_paper	False
vinli_simcse_v2	anti-ai/ViNLI-SimCSE-supervised_v2	mit	translated	True
virhe4qa	TaiDoan123/UIT-R2GQA (config triplet)	unknown	human_paper	False
viwikifc	High-Will/ViWikiFC	cc-by-4.0	human_paper	False

Citations

Please cite the method paper(s) first.

@misc{qu2021rocketqaoptimizedtrainingapproach,
      title={RocketQA: An Optimized Training Approach to Dense Passage Retrieval for Open-Domain Question Answering},
      author={Yingqi Qu and Yuchen Ding and Jing Liu and Kai Liu and Ruiyang Ren and Wayne Xin Zhao and Daxiang Dong and Hua Wu and Haifeng Wang},
      year={2021},
      eprint={2010.08191},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2010.08191},
}

@misc{thakur2025hardnegativeshardlessons,
      title={Hard Negatives, Hard Lessons: Revisiting Training Data Quality for Robust Information Retrieval with LLMs},
      author={Nandan Thakur and Crystina Zhang and Xueguang Ma and Jimmy Lin},
      year={2025},
      eprint={2505.16967},
      archivePrefix={arXiv},
      primaryClass={cs.IR},
      url={https://arxiv.org/abs/2505.16967},
}

@misc{wang2024multilinguale5textembeddings,
      title={Multilingual E5 Text Embeddings: A Technical Report},
      author={Liang Wang and Nan Yang and Xiaolong Huang and Linjun Yang and Rangan Majumder and Furu Wei},
      year={2024},
      eprint={2402.05672},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2402.05672},
}

@misc{chen2025m3embeddingmultilingualitymultifunctionalitymultigranularity,
      title={M3-Embedding: Multi-Linguality, Multi-Functionality, Multi-Granularity Text Embeddings Through Self-Knowledge Distillation},
      author={Jianlv Chen and Shitao Xiao and Peitian Zhang and Kun Luo and Defu Lian and Zheng Liu},
      year={2025},
      eprint={2402.03216},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2402.03216},
}

@inproceedings{csconda,
  title     = {A Benchmark Dataset and Evaluation Framework for Vietnamese Large Language Models in Customer Support},
  author    = {Nguyen, Long S. T. and Hua, Truong P. and Nguyen, Thanh M. and Pham, Toan Q. and Ngo, Nam K. and Nguyen, An X. and Pham, Nghi D. M. and Nguyen, Nghia H. and Quan, Tho T.},
  editor    = {Nguyen, Ngoc Thanh and Dinh Duc Anh, Vu and Kozierkiewicz, Adrianna and Nguyen Van, Sinh and Nunez, Manuel and Treur, Jan and Vossen, Gottfried},
  booktitle = {Advances in Computational Collective Intelligence},
  year      = {2026},
  publisher = {Springer Nature Switzerland},
  address   = {Cham},
  pages     = {481--496},
  isbn      = {978-3-032-10202-7}
}

@inproceedings{vire,
    title = "Which Works Best for {V}ietnamese? {A} Practical Study of Information Retrieval Methods across Domains",
    author = "Nguyen, Long S. T.  and
      Quan, Tho T.",
    editor = "Demberg, Vera  and
      Inui, Kentaro  and
      Marquez, Llu{\'i}s",
    booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EACL} 2026",
    month = mar,
    year = "2026",
    address = "Rabat, Morocco",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2026.findings-eacl.110/",
    doi = "10.18653/v1/2026.findings-eacl.110",
    pages = "2098--2119",
    ISBN = "979-8-89176-386-9"
}

@InProceedings{large-vi-legal-queries,
author="Son, Pham Tien and Hieu, Nguyen Doan and An, Nguyen Dai and Sang, Dinh Viet",
editor="Buntine, Wray and Fjeld, Morten and Tran, Truyen and Tran, Minh-Triet and Huynh Thi Thanh, Binh and Miyoshi, Takumi",
title="Improving Vietnamese Legal Document Retrieval Using Synthetic Data",
booktitle="Information and Communication Technology",
year="2025",
publisher="Springer Nature Singapore",
address="Singapore",
pages="378--393",
isbn="978-981-96-4291-5"
}

@misc{bonifacio2022mmarcomultilingualversionms,
      title={mMARCO: A Multilingual Version of the MS MARCO Passage Ranking Dataset}, 
      author={Luiz Bonifacio and Vitor Jeronymo and Hugo Queiroz Abonizio and Israel Campiotti and Marzieh Fadaee and Roberto Lotufo and Rodrigo Nogueira},
      year={2022},
      eprint={2108.13897},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2108.13897}, 
}

@misc{bajaj2018msmarcohumangenerated,
      title={MS MARCO: A Human Generated MAchine Reading COmprehension Dataset}, 
      author={Payal Bajaj and Daniel Campos and Nick Craswell and Li Deng and Jianfeng Gao and Xiaodong Liu and Rangan Majumder and Andrew McNamara and Bhaskar Mitra and Tri Nguyen and Mir Rosenberg and Xia Song and Alina Stoica and Saurabh Tiwary and Tong Wang},
      year={2018},
      eprint={1611.09268},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/1611.09268}, 
}

@misc{smart2026multiwikiqareadingcomprehensionbenchmark,
      title={MultiWikiQA: A Reading Comprehension Benchmark in 300+ Languages}, 
      author={Dan Saattrup Smart},
      year={2026},
      eprint={2509.04111},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2509.04111}, 
}

@inproceedings{gn-trvn,
  abstract = {Information retrieval often comes in plain text, lacking semi-structured text such as HTML and markdown, retrieving data that contains rich format such as table became non-trivial. In this paper, we tackle this challenge by introducing a new dataset, GreenNode Table Retrieval VN (GN-TRVN), which is collected from a massive corpus, a wide range of topics, and a longer context compared to ViQuAD2.0. To evaluate the effectiveness of our proposed dataset, we introduce two versions, M3-GN-VN and M3-GN-VN-Mixed, by fine-tuning the M3-Embedding model on this dataset. Experimental results show that our models consistently outperform the baselines, including the base model, across most evaluation criteria on various datasets such as VieQuADRetrieval, ZacLegalTextRetrieval, and GN-TRVN. In general, we release a more comprehensive dataset and two model versions that improve response performance for Vietnamese Markdown Table Retrieval.},
  address = {Singapore},
  author = {Pham, Bao Loc
and Hoang, Quoc Viet
and Luu, Quy Tung
and Vo, Trong Thu},
  booktitle = {Proceedings of the Fifth International Conference on Intelligent Systems and Networks},
  isbn = {978-981-95-1746-6},
  pages = {153--163},
  publisher = {Springer Nature Singapore},
  title = {GN-TRVN: A Benchmark for Vietnamese Table Markdown Retrieval Task},
  year = {2026},
}

@article{vianli,
title = {A New Benchmark Dataset and Mixture-of-Experts Language Models for Adversarial Natural Language Inference in Vietnamese},
journal = {Expert Systems with Applications},
pages = {130109},
year = {2025},
issn = {0957-4174},
doi = {https://doi.org/10.1016/j.eswa.2025.130109},
url = {https://www.sciencedirect.com/science/article/pii/S095741742503725X},
author = {Tin Van Huynh, Kiet Van Nguyen and Ngan Luu-Thuy Nguyen},
}

@inproceedings{thai-etal-2022-uit,
    title = "{UIT}-{V}i{C}o{V}19{QA}: A Dataset for {COVID}-19 Community-based Question Answering on {V}ietnamese Language",
    author = "Thai, Triet  and
      Thao-Ha, Ngan Chu  and
      Vo, Anh  and
      Luu, Son",
    editor = "Dita, Shirley  and
      Trillanes, Arlene  and
      Lucas, Rochelle Irene",
    booktitle = "Proceedings of the 36th Pacific {A}sia Conference on Language, Information and Computation",
    month = oct,
    year = "2022",
    address = "Manila, Philippines",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2022.paclic-1.88/",
    pages = "801--810"
}

@inproceedings{tran2025vifactcheck,
  title     = {ViFactCheck: A New Benchmark Dataset and Methods for Multi-Domain News Fact-Checking in Vietnamese},
  author    = {Tran, Thai Hoa and Tran, Quang Duy and Tran, Khanh Quoc and Nguyen, Kiet Van},
  booktitle = {Proceedings of the Thirty-Ninth AAAI Conference on Artificial Intelligence (AAAI-25)},
  pages     = {308--316},
  year      = {2025},
  publisher = {AAAI Press}
}

@inproceedings{tran-etal-2024-vimedaqa,
    title = "{V}i{M}ed{AQA}: A {V}ietnamese Medical Abstractive Question-Answering Dataset and Findings of Large Language Model",
    author = "Tran, Minh-Nam  and
      Nguyen, Phu-Vinh  and
      Nguyen, Long  and
      Dinh, Dien",
    editor = "Fu, Xiyan  and
      Fleisig, Eve",
    booktitle = "Proceedings of the 62nd Annual Meeting of the {A}ssociation for {C}omputational {L}inguistics (Volume 4: Student Research Workshop)",
    month = aug,
    year = "2024",
    address = "Bangkok, Thailand",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2024.acl-srw.31/",
    doi = "10.18653/v1/2024.acl-srw.31",
    pages = "252--260",
    ISBN = "979-8-89176-097-4"
}

@inproceedings{le-etal-2022-vimqa,
    title = "{VIMQA}: A {V}ietnamese Dataset for Advanced Reasoning and Explainable Multi-hop Question Answering",
    author = "Le, Nguyen-Khang  and
      Nguyen, Dieu-Hien  and
      Le, Tung  and
      Nguyen, Minh",
    editor = "Calzolari, Nicoletta  and
      B{\'e}chet, Fr{\'e}d{\'e}ric  and
      Blache, Philippe  and
      Choukri, Khalid  and
      Cieri, Christopher  and
      Declerck, Thierry  and
      Goggi, Sara  and
      Isahara, Hitoshi  and
      Maegaard, Bente  and
      Mariani, Joseph  and
      Mazo, H{\'e}l{\`e}ne  and
      Odijk, Jan  and
      Piperidis, Stelios",
    booktitle = "Proceedings of the Thirteenth Language Resources and Evaluation Conference",
    month = jun,
    year = "2022",
    address = "Marseille, France",
    publisher = "European Language Resources Association",
    url = "https://aclanthology.org/2022.lrec-1.700/",
    pages = "6521--6529"
}

@inproceedings{huynh-etal-2022-vinli,
    title = "{V}i{NLI}: A {V}ietnamese Corpus for Studies on Open-Domain Natural Language Inference",
    author = "Huynh, Tin Van  and
      Nguyen, Kiet Van  and
      Nguyen, Ngan Luu-Thuy",
    editor = "Calzolari, Nicoletta  and
      Huang, Chu-Ren  and
      Kim, Hansaem  and
      Pustejovsky, James  and
      Wanner, Leo  and
      Choi, Key-Sun  and
      Ryu, Pum-Mo  and
      Chen, Hsin-Hsi  and
      Donatelli, Lucia  and
      Ji, Heng  and
      Kurohashi, Sadao  and
      Paggio, Patrizia  and
      Xue, Nianwen  and
      Kim, Seokhwan  and
      Hahm, Younggyun  and
      He, Zhong  and
      Lee, Tony Kyungil  and
      Santus, Enrico  and
      Bond, Francis  and
      Na, Seung-Hoon",
    booktitle = "Proceedings of the 29th International Conference on Computational Linguistics",
    month = oct,
    year = "2022",
    address = "Gyeongju, Republic of Korea",
    publisher = "International Committee on Computational Linguistics",
    url = "https://aclanthology.org/2022.coling-1.339/",
    pages = "3858--3872"
}

@inproceedings{williams-etal-2018-broad,
    title = "A Broad-Coverage Challenge Corpus for Sentence Understanding through Inference",
    author = "Williams, Adina  and
      Nangia, Nikita  and
      Bowman, Samuel R.",
    editor = "Walker, Marilyn  and
      Ji, Heng  and
      Stent, Amanda",
    booktitle = "Proceedings of the 2018 Conference of the {N}orth {A}merican Chapter of the {A}ssociation for {C}omputational {L}inguistics: Human Language Technologies, Volume 1 (Long Papers)",
    month = jun,
    year = "2018",
    address = "New Orleans, Louisiana",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/N18-1101/",
    doi = "10.18653/v1/N18-1101",
    pages = "1112--1122"
}

@misc{do2024r2gqaretrieverreadergeneratorquestionanswering,
      title={R2GQA: Retriever-Reader-Generator Question Answering System to Support Students Understanding Legal Regulations in Higher Education}, 
      author={Phuc-Tinh Pham Do and Duy-Ngoc Dinh Cao and Khanh Quoc Tran and Kiet Van Nguyen},
      year={2024},
      eprint={2409.02840},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2409.02840}, 
}

@inproceedings{le-etal-2026-viwikifc,
    title = "{V}i{W}iki{FC}: Fact-Checking for {V}ietnamese {W}ikipedia-Based Textual Knowledge Source",
    author = "Le, Hung Tuan  and
      To, Long Truong  and
      Nguyen, Manh Trong  and
      Nguyen, Kiet Van",
    editor = "Piperidis, Stelios  and
      Bel, N{\'u}ria  and
      van den Heuvel, Henk  and
      Ide, Nancy  and
      Krek, Simon  and
      Toral, Antonio",
    booktitle = "Proceedings of the Fifteenth Language Resources and Evaluation Conference",
    month = may,
    year = "2026",
    address = "Palma de Mallorca, Spain",
    publisher = "ELRA Language Resource Association",
    url = "https://aclanthology.org/2026.lrec-1.869/",
    doi = "10.63317/5ew98tfds6jc",
    pages = "11125--11140"
}

No verified citation was found for: isedsc2024, vietnamese_law_hn_mined.

Limitations and licenses

Each source keeps its own license (column above); this collection is license: other. Sources that are non-commercial or have no stated license: csconda, isedsc2024, multi_wiki_qa, vicov19qa, vietnamese_law_hn_mined, vihealthqa, vimedaqa, vimqa, vinli, virhe4qa. Check the original license before any use beyond research. Q&A sources crawled from the web had person names masked with a Vietnamese NER model and phone/ID/account numbers masked by regex (<name>, <phone>, …).