Datasets:

aryopg
/

parametric-arithmetic-eval

license: mit
language:
  - en
task_categories:
  - question-answering
pretty_name: Parametric & Arithmetic Eval
size_categories:
  - n<1K
tags:
  - retrieval-heads
  - mechanistic-interpretability
  - parametric-knowledge
  - ablation
dataset_info:
  features:
    - name: question
      dtype: string
    - name: answer
      dtype: string
    - name: source
      dtype: string
    - name: index
      dtype: int64
  splits:
    - name: train
      num_bytes: 40056
      num_examples: 600
  download_size: 16223
  dataset_size: 40056
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*

Parametric & Arithmetic Eval

A 600-example control set for testing whether ablated attention heads are retrieval-specific rather than generically important for model output. Every question is answerable from the model's own parametric knowledge or by direct computation — none require retrieving information from an in-context document.

This dataset accompanies LOCOS (Logit-Contribution Scoring). A retrieval-head detector is only meaningful if ablating the heads it identifies degrades contextual retrieval while leaving non-retrieval abilities intact.

Composition

600 examples, 200 per source:

`source`	Count	Task	Example question	Example answer
`city_country`	200	Parametric factual recall	Which country does Tokyo belong to?	`Japan`
`popqa`	200	Parametric factual recall	Who is the father of Brandon Lee?	`Bruce Lee`
`arithmetic`	200	Two-operand add/subtract	What is (5 + 6) - 3?	`8`

Fields

Field	Type	Description
`question`	string	The prompt question.
`answer`	string	Reference answer (short form).
`source`	string	One of `city_country`, `popqa`, `arithmetic`.
`index`	int64	Global row index (0–199 city_country, 200–399 popqa, 400–599 arithmetic).

Provenance

city_country — top-200 cities by population from WorkWithData/cities, reformatted as country-identification questions (country names normalised to common English forms).
popqa — most-popular entries from akariasai/PopQA, ranked by the geometric mean of subject/object popularity and deduplicated by subject–object pair.
arithmetic — random sample (seed 42) from EleutherAI/arithmetic (arithmetic_1dc), with the Question:/Answer: scaffolding stripped.

Usage

from datasets import load_dataset

ds = load_dataset("aryopg/parametric-arithmetic-eval", split="train")
city_country = ds.filter(lambda r: r["source"] == "city_country")

License

Released under the MIT License. The dataset is derived from openly available sources; please also respect the licenses of the upstream datasets linked above.

Citation

If you use this dataset, please cite LOCOS and PopQA:

@article{gema2026locos,
      title={Logit-Contribution Scoring Identifies Non-Literal Retrieval Heads}, 
      author={Aryo Pradipta Gema and Beatrice Alex and Pasquale Minervini},
      year={2026},
      eprint={2607.01002},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2607.01002}, 
}

@article{mallen2023llm_memorization,
  title   = {When Not to Trust Language Models: Investigating Effectiveness and Limitations of Parametric and Non-Parametric Memories},
  author  = {Mallen, Alex and Asai, Akari and Zhong, Victor and Das, Rajarshi and Hajishirzi, Hannaneh and Khashabi, Daniel},
  journal = {arXiv preprint},
  year    = {2022}
}

This dataset also benefits from:

City-Country: WorkWithData/cities
Arithmetic: EleutherAI/arithmetic, subset arithmetic_1dc

Note: the LOCOS citation is a placeholder pending the arXiv release.