Datasets:

AlgorithmicResearchGroup
/

aria-search-benchmark_v2-public

pretty_name: ARIA Search Benchmark v2
dataset_info:
  features:
    - name: dataset
      dtype: string
    - name: model_name
      dtype: string
    - name: paper_title
      dtype: string
    - name: paper_date
      dtype: timestamp[ns]
    - name: paper_url
      dtype: string
    - name: code_links
      list: string
    - name: prompts
      dtype: string
    - name: answer
      dtype: string
    - name: paper_text
      dtype: string
    - name: year_bin
      dtype: string
    - name: benchmark_split
      dtype: string
  splits:
    - name: benchmark
      num_bytes: 223990320
      num_examples: 3517
  download_size: 54514058
  dataset_size: 223990320
configs:
  - config_name: default
    data_files:
      - split: benchmark
        path: data/benchmark-*
language:
  - en
size_categories:
  - 1K<n<10K
tags:
  - aria
  - benchmark
  - ml-research
  - search
  - retrieval
task_categories:
  - question-answering

ARIA Search Benchmark v2

The ARIA Search Benchmark is part of the ARIA benchmark suite, a collection of closed-book benchmarks probing the ML knowledge that frontier models have internalized during training. This dataset tests whether models can answer factual questions about ML research papers, models, datasets, and benchmark results without access to external retrieval.

Dataset Summary

Size: 3,517 question-answer pairs
Split: benchmark
Paper date range: May 2023 to December 2024
Coverage: Spans models, datasets, and metrics across CV, NLP, audio, video, and multimodal domains

Dataset Structure

Field	Type	Description
`dataset`	string	Benchmark dataset referenced
`model_name`	string	Model being evaluated
`paper_title`	string	Source paper title
`paper_date`	timestamp	Publication date
`paper_url`	string	ArXiv paper URL
`code_links`	list[string]	GitHub repository links
`prompts`	string	Question/prompt text
`answer`	string	Ground-truth answer
`paper_text`	string	Full paper text
`year_bin`	string	Year category for stratified evaluation
`benchmark_split`	string	Benchmark split identifier

Usage

from datasets import load_dataset

ds = load_dataset("AlgorithmicResearchGroup/aria-search-benchmark_v2-public", split="benchmark")

for example in ds.select(range(5)):
    print(f"Q: {example['prompts']}")
    print(f"A: {example['answer']}")
    print(f"Paper: {example['paper_title']}")
    print()

Related Resources

Citation

@misc{aria_search_benchmark_v2,
    title={ARIA Search Benchmark v2},
    author={Algorithmic Research Group},
    year={2024},
    publisher={Hugging Face},
    url={https://huggingface.co/datasets/AlgorithmicResearchGroup/aria-search-benchmark_v2-public}
}