SIG_Wiki

SIG_Wiki (Spurious features In Golden documents from Wiki) is a lightweight benchmark for evaluating the robustness of Retrieval-Augmented Language Models (RALMs) against spurious features — semantic-agnostic modifications to grounding documents that should not change the answer, but often do.

It is the challenge subset of the SURE_Wiki dataset from the paper Quantifying and Improving the Robustness of Retrieval-Augmented Language Models Against Spurious Features in Grounding Data (ACL 2026). Each entry is a hard case where both Mistral-7B-Instruct-v0.3 and Llama-3.1-8B-Instruct were non-robust. Queries come from NQ-open; documents come from a Wikipedia dump (retrieved with Contriever-msmarco).

Code: https://github.com/maybenotime/RAG-SpuriousFeatures

Structure

1,600 instances = 5 feature types × their perturbations × 100 instances each.

feature_type	perturbation values	rows
format	format_preference_html, _json, _markdown, _yaml	400
style	simple_preference_llm, complex_preference_llm	200
logic	logic_preference_reverse, _random, _llm	300
source	source_preference_llm, self_preference_mistral_7b, self_preference_llama3_8b	300
meta	metadata_timestamp_pre, _post, metadata_datasource_wiki, _twitter	400

Fields

• feature_type — spurious feature category.
• perturbation — the specific perturbation applied.
• question — the query (NQ-open).
• answers — list of acceptable gold answers.
• title — document title.
• original_text — the un-perturbed golden document.
• perturbed_doc — the same document after the spurious feature is injected.

The raw nested JSON (grouped by feature type / perturbation) is also available at raw/SIG_benchmark.json.

Usage

from datasets import load_dataset

ds = load_dataset("maybenotime/SIG_Wiki", split="test")
fmt = ds.filter(lambda r: r["feature_type"] == "format")

Feed original_text and perturbed_doc (with the same question) to a RALM and compare the correctness of the two answers. Following the paper, report the instance-level Robustness Rate / Win Rate / Lose Rate.

Citation

@inproceedings{yang2026sure,
  title     = {Quantifying and Improving the Robustness of Retrieval-Augmented Language Models Against Spurious Features in Grounding Data},
  author    = {Yang, Shiping and Wu, Jie and Ding, Wenbiao and Wu, Ning and Liang, Shining and Gong, Ming and Li, Hongzhi and Zhang, Hengyuan and Chang, Angel X. and Zhang, Dongmei},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  year      = {2026},
  pages     = {33479--33499}
}