lab-flair/qa-dataset-k1000

QA Dataset K1000 — The First Drop of Ink

Question-answering data with gold documents and distractor pools for long-context evaluation, accompanying The First Drop of Ink: Nonlinear Impact of Distracting Information in Long-Context Reasoning by Muhan Gao, Zih-Ching Chen, and Kuan-Hao Huang (ICML 2026).

Paper · Full text (v2) · Hugging Face paper page

The paper studies how the proportion of hard distractors affects performance at fixed context length. It reports a nonlinear decline: the first small fraction of hard distractors causes a disproportionately large drop. See the paper for context construction, model settings, evaluation, and limitations; downloading these files alone does not reproduce the full experimental protocol.

Files

File	Source question-answering task
nq_k1000.jsonl	Natural Questions
popqa_k1000.jsonl	PopQA
triviaqa_k1000.jsonl	TriviaQA

Each JSONL file contains one question-answering example per line. The K1000 files provide pools of 1,000 distractor documents per example; the experiment's final context is constructed from these pools. The Hub currently indexes 904 examples in its automatically generated train split. That split name is a Hub loading convention, not a recommendation to train on evaluation examples.

Data format

Core fields:

Field	Description
question	Question text.
answer	List of accepted answer strings.
gold_docs	Gold-document records.
distractors	Distractor-document records.

Document records include id, title, text, score, and has_answer. Additional document collections, where present, are named D_rel, D_modal, D_hypo, and D_neg; retain these fields when loading the data. Refer to the paper's experimental setup when selecting or mixing distractors.

Download and load

Install the download client:

pip install huggingface_hub

Download and inspect one JSONL file:

import json
from huggingface_hub import hf_hub_download

file_path = hf_hub_download(
    repo_id="lab-flair/qa-dataset-k1000",
    filename="nq_k1000.jsonl",
    repo_type="dataset",
)

with open(file_path, encoding="utf-8") as f:
    example = json.loads(next(f))

print(example["question"])
print(example["answer"])
print(len(example["distractors"]))

Use popqa_k1000.jsonl or triviaqa_k1000.jsonl to download the other tasks. To make a download reproducible, add revision="<dataset-commit-sha>" to hf_hub_download, using a commit from this repository's Files and versions history.

The canonical repository ID is lab-flair/qa-dataset-k1000.

Citation and provenance

If you use this release to study the distractor effect, or build on the analysis and experimental protocol in The First Drop of Ink, please cite the paper:

@misc{gao2026firstdropink,
  title = {The First Drop of Ink: Nonlinear Impact of Distracting Information in Long-Context Reasoning},
  author = {Muhan Gao and Zih-Ching Chen and Kuan-Hao Huang},
  year = {2026},
  eprint = {2605.10828},
  archivePrefix = {arXiv},
  primaryClass = {cs.AI},
  note = {Version 2, August 20, 2026},
  url = {https://arxiv.org/abs/2605.10828v2}
}

This release also retains the original dataset card's attribution to RULER. Cite RULER when using its benchmark setup or derived resources, and credit the underlying Natural Questions, PopQA, and TriviaQA datasets as applicable. The First Drop of Ink citation does not replace upstream attribution.

The following RULER citation is provided by its official repository:

@article{hsieh2024ruler,
  title = {RULER: What's the Real Context Size of Your Long-Context Language Models?},
  author = {Cheng-Ping Hsieh and Simeng Sun and Samuel Kriman and Shantanu Acharya and Dima Rekesh and Fei Jia and Yang Zhang and Boris Ginsburg},
  year = {2024},
  journal = {arXiv preprint arXiv:2404.06654}
}