Datasets:

styme3279
/

control-apps-cleaned

license: mit
task_categories:
  - text-generation
language:
  - en
tags:
  - code
  - python
  - ai-safety
  - control
  - apps
pretty_name: APPS-cleaned for AI Control replication
size_categories:
  - 1K<n<10K

control-apps-cleaned

A curated subset of the APPS dataset (Hendrycks et al., NeurIPS 2021), pre-filtered for use in the ARENA AI Control chapter — a teaching replication of Greenblatt et al. 2023 (arXiv:2312.06942).

What's in here

cleaned_apps.jsonl — 1,202 problems from the APPS "interview" split, filtered to a uniform I/O schema (inputs and outputs are each one of list[str], list[int], list[list[str]], list[list[int]]). Each line is a JSON record with the original APPS fields plus two added fields:
- input_type — one of "list of strings", "list of integers", "list of lists of strings", "list of lists of integers"
- output_type — same vocabulary

The filtering removes problems whose original I/O was strings-with-embedded-newlines (which broke the chapter's main(input) function-call interface).

Why this exists

The original APPS dataset has inputs as raw stdin strings, which doesn't play nicely with the inspect_ai sandbox calling main(input) directly. This filtered subset lets every problem be solved by a function def main(input: list[str]) -> list[str]: ... without per-problem dispatch on the input shape.

Usage

from huggingface_hub import hf_hub_download
import json

path = hf_hub_download(
    repo_id="styme3279/control-apps-cleaned",
    filename="cleaned_apps.jsonl",
    repo_type="dataset",
)

with open(path) as f:
    apps = [json.loads(line) for line in f]

print(len(apps), "problems")
print(apps[0].keys())

Provenance

Upstream source: hendrycks/apps, interview-difficulty split
Filter script: utils.clean_dataset in the ARENA chapter chapter3_llm_evals/exercises/part5_ai_control/utils.py
Used by: ARENA chapter 3.5 (AI Control)

License

MIT, inherited from the upstream APPS dataset.