pauljngr/sardi-data

SARDI — Evaluation Data

Test splits and prebuilt BM25 indices for Self-Augmenting Retrieval for Diffusion Language Models (ICML 2026).

Paper · Code · Model

Download

hf download pauljngr/sardi-data --repo-type dataset --local-dir data

Contents

dataset	questions	passages	size
2WikiMultiHopQA	6,253	406,822	308 MB
HotpotQA	3,701	5,239,002	2.7 GB
MuSiQue	2,417	103,035	92 MB
CofCA	900	3,156	6 MB
SynthWorlds-SM	1,200	8,055	16 MB

Size is the whole dataset: the test split plus the index directory, which carries its own copy of the corpus. All five together are 3.1 GB.

<dataset>/test.parquet                   question, answer, supporting_docs
<dataset>/corpus/index_chunked/          bm25s index + its copy of the corpus
manifest.json                            sha256 of every file, row counts

Passages are ~300-token chunks with one sentence of overlap. For HotpotQA we retrieve from their full Wikipedia corpus; for the other four we follow prior work, e.g., DEXTER: A Benchmark for Open-domain Complex Question Answering using LLMs (Venktesh V et al., 2024) and Unlocking the Potentials of Retrieval-Augmented Generation for Diffusion Language Models (Yu et al., 2026), and concatenate all gold and distractor passages into a single corpus. This way, the corpora end up at the same granularity as each dataset's gold-document annotations, making it possible to measure retrieval recall.

Counterfactual benchmarks

CofCA and SynthWorlds-SM use corpora of invented facts, absent from both pre-training and our fine-tuning data, so they measure retrieval rather than memorization.

Licensing

These are reformatted and chunked versions of publicly released benchmarks. Each retains its original license.

2WikiMultiHopQA	Constructing A Multi-hop QA Dataset for Comprehensive Evaluation of Reasoning Steps — Ho et al., COLING 2020
HotpotQA	HotpotQA: A Dataset for Diverse, Explainable Multi-hop Question Answering — Yang et al., EMNLP 2018
MuSiQue	MuSiQue: Multihop Questions via Single-hop Question Composition — Trivedi et al., TACL 2022
CofCA	CofCA: A Step-Wise Counterfactual Multi-hop QA Benchmark — Wu et al., ICLR 2025
SynthWorlds	SynthWorlds: Controlled Parallel Worlds for Disentangling Reasoning and Knowledge in Language Models — Gu et al., 2025

Citation

If you use these splits in your work, please cite the original benchmark papers above, alongside:

@inproceedings{juenger2026sardi,
  title     = {Self-Augmenting Retrieval for Diffusion Language Models},
  author    = {Paul J{\"u}nger and Justin Lovelace and Linxi Zhao and Dongyoung Go and Kilian Q Weinberger},
  booktitle = {International Conference on Machine Learning (ICML)},
  year      = {2026}
}