metadata
license: apache-2.0
configs:
- config_name: agnews
data_files: agnews.jsonl.tar.gz
- config_name: AllNLI
data_files: AllNLI.jsonl.tar.gz
- config_name: altlex
data_files: altlex.jsonl.tar.gz
- config_name: amazon_review_2018_1m
data_files: amazon_review_2018_1m.jsonl.tar.gz
- config_name: cnn_dailymail
data_files: cnn_dailymail.jsonl.tar.gz
- config_name: codesearchnet
data_files: codesearchnet.jsonl.tar.gz
- config_name: dureader
data_files: dureader.jsonl.tar.gz
- config_name: eli5_question_answer
data_files: eli5_question_answer.jsonl.tar.gz
- config_name: gooaq_pairs
data_files: gooaq_pairs.jsonl.tar.gz
- config_name: hotpotqa
data_files: hotpotqa.jsonl.tar.gz
- config_name: medmcqa
data_files: medmcqa.jsonl.tar.gz
- config_name: miracl
data_files: miracl.jsonl.tar.gz
- config_name: mr_tydi_combined
data_files: mr_tydi_combined.jsonl.tar.gz
- config_name: msmarco
data_files: msmarco.jsonl.tar.gz
- config_name: nq
data_files: nq.jsonl.tar.gz
- config_name: quora_duplicates_triplets
data_files: quora_duplicates_triplets.jsonl.tar.gz
- config_name: searchQA_top5_snippets
data_files: searchQA_top5_snippets.jsonl.tar.gz
- config_name: sentence-compression
data_files: sentence-compression.jsonl.tar.gz
- config_name: SimpleWiki
data_files: SimpleWiki.jsonl.tar.gz
- config_name: squad_pairs
data_files: squad_pairs.jsonl.tar.gz
- config_name: stackexchange_duplicate_questions_title-body_title-body
data_files: stackexchange_duplicate_questions_title-body_title-body.jsonl.tar.gz
- config_name: t2ranking
data_files: t2ranking.jsonl.tar.gz
- config_name: trivia
data_files: trivia.jsonl.tar.gz
- config_name: xsum
data_files: xsum.jsonl.tar.gz
- config_name: yahoo_answers_title_answer
data_files: yahoo_answers_title_answer.jsonl.tar.gz
tdro-llm/finetune_data
tDRO: Task-level Distributionally Robust Optimization for Large Language Model-based Dense Retrieval. Guangyuan Ma, Yongliang Ma, Xing Wu, Zhenpeng Su, Ming Zhou and Songlin Hu.
This repo contains all fine-tuning data for Large Language Model-based Dense Retrieval. Please refer to this repo for details to reproduce.
A total of 25 heterogeneous retrieval fine-tuning datasets with Hard Negatives and Deduplication (with test sets) are listed as belows.
| Dataset | Language | Category | Symmetry | Reference | Format | HN Mine | Size | Deduped Size | Duplicates |
|---|---|---|---|---|---|---|---|---|---|
| agnews | English | News | Asymmetric | AG news corpus | (Title, Description, Negatives) of news articles from the AG News dataset | bge-base-en-v1.5 mine | 1157745 | 1157745 | 0 |
| AllNLI | English | NLI | Symmetric | SNLI and MNLI | (Anchor, Entailment_Text, Contradiction_Text) - Combination of SNLI + MultiNLI Triplets | Sentence Transformers Train HN | 277230 | 277230 | 0 |
| altlex | English | Wikipedia Pair | Symmetric | altlex | (English_Wikipedia, Simple_English_Wikipedia, Negatives) - Matched pairs | bge-base-en-v1.5 mine | 112696 | 112696 | 0 |
| amazon_review_2018_1m | English | Amazon | Asymmetric | Amazon review data (2018) | (Title, review, Negatives) from Amazon. Only Top 1 million samples are used here to shrink the dataset sizes. | bge-base-en-v1.5 mine | 1000000 | 999999 | 1 |
| cnn_dailymail | English | News | Asymmetric | CNN Dailymail Dataset | (highlight sentences, article) with all highlight sentences as one text for each news article | bge-base-en-v1.5 mine | 311971 | 311971 | 0 |
| codesearchnet | English | Github | Asymmetric | CodeSearchNet | (Comment, Code, Negatives) - pairs from opensource libraries hosted on GitHub. It contains code and documentation for several programming languages. | bge-base-en-v1.5 mine | 1375067 | 1375067 | 0 |
| dureader | Chinese | Multilingual Web Collections | Asymmetric | DuReader Retrieval | (Question, Answer, Negatives) | Use originally provided HN | 86395 | 86395 | 0 |
| eli5_question_answer | English | Asymmetric | ELI5 | (Question, Answer, Negatives) | bge-base-en-v1.5 mine | 325475 | 325390 | 85 | |
| gooaq_pairs | English | Web Collections | Asymmetric | GooAQ | (Question, Answer, Negatives) - Pairs from Google auto suggest | bge-base-en-v1.5 mine | 3012496 | 3012347 | 149 |
| hotpotqa | English | Wikipedia QA | Asymmetric | HotpotQA | (Question, Answer, Negatives) | bge-base-en-v1.5 mine | 85000 | 85000 | 0 |
| medmcqa | English | Medical | Asymmetric | MedMCQA | (Question, Answer, Negatives) | bge-base-en-v1.5 mine | 160869 | 160865 | 4 |
| miracl | 16 languages | Multilingual Wikipedia | Asymmetric | MIRACL | (Question, Answer, Negatives) | Use originally provided HN | 32561 | 32405 | 156 |
| mr_tydi_combined | 11 languages | Multilingual Wikipedia | Asymmetric | Mr. TyDi | (Question, Answer, Negatives) | Use originally provided HN | 48715 | 48475 | 240 |
| msmarco | English | Web Collections | Asymmetric | MS MARCO Passages | (Question, Answer, Negatives) | bowdpr HN by following this link | 502939 | 502854 | 85 |
| nq | English | Wikipedia QA | Asymmetric | NQ | (Question, Answer, Negatives) | bowdpr HN by following this link | 58812 | 58800 | 12 |
| quora_duplicates_triplets | English | Forum Duplicates | Symmetric | QQP | (Question, Duplicated Question, Negatives) - Duplicate question pairs from Quora | Sentence Transformers Train HN | 101762 | 97011 | 4751 |
| searchQA_top5_snippets | English | Web Collections | Asymmetric | search_qa | (Question, Top5 text snippets, Negatives) from SearchQA dataset. | bge-base-en-v1.5 mine | 117220 | 117219 | 1 |
| sentence-compression | English | News | Asymmetric | Sentence-Compression | (Long text, short text) about sentence-compression | bge-base-en-v1.5 mine | 180000 | 180000 | 0 |
| SimpleWiki | English | Wikipedia Pair | Symmetric | SimpleWiki | (English_Wikipedia, Simple_English_Wikipedia, Negatives) matched pairs | bge-base-en-v1.5 mine | 102225 | 102225 | 0 |
| squad_pairs | English | Wikipedia QA | Asymmetric | SQuAD | (Question, Answer, Negatives) | bge-base-en-v1.5 mine | 87599 | 87595 | 4 |
| stackexchange_duplicate_questions_title-body_title-body | English | Forum Duplicates | Symmetric | Stack Exchange Data API | (Title-Body, Duplicated Title-Body, Negatives) - pairs of duplicate questions from StackExchange | bge-base-en-v1.5 mine | 250519 | 250516 | 3 |
| t2ranking | Chinese | Multilingual Web Collections | Asymmetric | T2Ranking | (Question, Answer, Negatives) | Use originally provided HN | 200376 | 200376 | 0 |
| trivia | English | Wikipedia QA | Asymmetric | Trivia QA | (Question, Answer, Negatives) | bowdpr HN by following this link | 60380 | 60370 | 10 |
| xsum | English | News | Asymmetric | xsum | (Summary, News Article) pairs from XSUM dataset | bge-base-en-v1.5 mine | 226711 | 226711 | 0 |
| yahoo_answers_title_answer | English | Yahoo | Asymmetric | Yahoo Answers | (Title, Answer, Negatives) | bge-base-en-v1.5 mine | 1198260 | 1198018 | 242 |
| Total Lines | 11073023 | 11067280 | 5743 |
Cite
If you are interested in our work, please consider citing our paper.
@article{ma2024tdro,
author = {Guangyuan Ma and
Yongliang Ma and
Xing Wu and
Zhenpeng Su and
Ming Zhou and
Songlin Hu},
title = {Task-level Distributionally Robust Optimization for Large Language
Model-based Dense Retrieval},
journal = {CoRR},
volume = {abs/2408.10613},
year = {2024},
url = {https://doi.org/10.48550/arXiv.2408.10613},
doi = {10.48550/ARXIV.2408.10613},
eprinttype = {arXiv},
eprint = {2408.10613},
timestamp = {Tue, 24 Sep 2024 17:36:32 +0200},
}