OneWordQA

1,548 factual questions, each with a one-word answer, in 23 categories: capitals, chemical symbols, dialing codes, Roman numerals and the like. Each question on its own is easy for a current reasoning model. That makes it a clean probe for the paper On the Fragility of Tasks under Composition: Implicit Budgeting in Language Models, which asks many such questions in one prompt and measures how accuracy falls as their number grows. The code for those experiments is at github.com/chenpanli/fragility.

from datasets import load_dataset

pool = load_dataset("chenpanli/onewordqa", split="test")

How it was made

The questions and their answers were generated with Claude Opus 5 and then reviewed by hand. The reviewed list is generation/datasets/onewordqa/base.jsonl, one row per question with its category and every accepted spelling of the answer.

generation/experiments/onewordqa/generate_dataset.py turns that list into the released pool. It drops malformed rows and duplicate questions (none in this list), groups rows by category, assigns the stable q-NNNN ids, and renders each question as it appears in a prompt. It also writes onewordqa.manifest.json, which records the SHA-256 of base.jsonl and the category counts. To rebuild the pool:

cd generation
PYTHONPATH=src python experiments/onewordqa/generate_dataset.py --seed 42
cmp datasets/onewordqa/atomic.jsonl ../onewordqa.jsonl    # identical

The script needs only the Python 3.12 standard library. generation/src/llm/ holds the two modules it imports from the paper's code: onewordqa.py (category groups and the row check) and rendering.py (the question text, identical to the code repository's copy).

Fields

Field	Meaning
id	Stable id, q-0000 to q-1547. Every composed prompt in the paper refers to questions by this id.
question	The question text.
correct_answers	Every accepted spelling of the answer. Four questions accept two.
accepted_answers	[correct_answers], a list per question so a single question looks like a composition of one.
reference_answer, reference_answers	The canonical answer, alone and as a one-element list.
n_accepted	len(correct_answers).
questions_block	The question as rendered into the prompt (1. <question>).
category, broad_category	One of 23 categories, grouped into geography (687), science (419), math (269), everyday (113) and games and sport (60).
stage, k	Always atomic and 1.
source_dataset, source_id	The row of base.jsonl the question came from.

A sample row:

{"id": "q-0008", "question": "What is the capital city of Algeria?", "correct_answers": ["Algiers"],
 "category": "country_capital", "broad_category": "geography", "questions_block": "1. What is the capital city of Algeria?", ...}

Answers are graded by string match against correct_answers after both sides are normalised for how they are written: case, accents, punctuation, a leading article, and number words or ordinals (four and 4th count as 4). There are no aliases, so Mammalia does not match Mammal. The grader is src/llm/core/oneword/ in the code repository.

Citation

@inproceedings{li2026fragility,
  title     = {On the Fragility of Tasks under Composition: Implicit Budgeting in Language Models},
  author    = {Li, Chenpan and Brockmoeller, Thies and Jegelka, Stefanie and
               Gupta, Sharut and Karalias, Nikolaos},
  year      = {2026},
}