README.md

metadata

pretty_name: ARIA Repo Benchmark
dataset_info:
  features:
    - name: dataset
      dtype: string
    - name: model_name
      dtype: string
    - name: model_links
      list: 'null'
    - name: paper_title
      dtype: string
    - name: paper_date
      dtype: timestamp[ns]
    - name: paper_url
      dtype: string
    - name: code_links
      list: string
    - name: metrics
      dtype: string
    - name: table_metrics
      list: string
    - name: prompts
      dtype: string
    - name: paper_text
      dtype: string
    - name: compute_hours
      dtype: float64
    - name: num_gpus
      dtype: int64
    - name: reasoning
      dtype: string
    - name: trainable_single_gpu_8h
      dtype: string
    - name: verified
      dtype: string
    - name: modality
      dtype: string
    - name: paper_title_drop
      dtype: string
    - name: paper_date_drop
      dtype: string
    - name: code_links_drop
      dtype: string
    - name: num_gpus_drop
      dtype: int64
    - name: dataset_link
      dtype: string
    - name: time_and_compute_verification
      dtype: string
    - name: link_to_colab_notebook
      dtype: string
    - name: run_possible
      dtype: string
    - name: notes
      dtype: string
  splits:
    - name: train
      num_bytes: 3507790
      num_examples: 58
  download_size: 1553348
  dataset_size: 3507790
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
language:
  - en
size_categories:
  - n<1K
tags:
  - aria
  - benchmark
  - ml-research
  - reproducibility
task_categories:
  - text-generation

ARIA Repo Benchmark

The ARIA Repo 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 contains 58 curated research paper implementations with metadata for evaluating whether ML experiments described in papers can be reproduced.

Dataset Summary

• Size: 58 entries
• Coverage: Computer Vision, NLP, Time Series, Graph, Bioinformatics
• Purpose: Evaluate AI agents on their ability to locate, understand, and reproduce ML research experiments

Each entry links a research paper to its code repository, dataset, metrics, and compute requirements, along with verification of whether the experiment is reproducible on constrained hardware.

Dataset Structure

Key Fields

Field	Type	Description
paper_title	string	Title of the research paper
paper_url	string	ArXiv URL
paper_date	timestamp	Publication date
paper_text	string	Full paper text
dataset	string	Dataset used in the paper
dataset_link	string	Link to the dataset
model_name	string	Model name
code_links	list[string]	GitHub repository links
metrics	string	Performance metrics reported
table_metrics	list[string]	Detailed metrics from tables
prompts	string	Evaluation prompts
modality	string	Data modality (CV, NLP, Time Series, Graph, etc.)

Compute & Reproducibility Fields

Field	Type	Description
compute_hours	float64	Estimated training compute hours
num_gpus	int64	Number of GPUs required
reasoning	string	Reasoning about compute estimates
trainable_single_gpu_8h	string	Trainable on a single GPU in 8 hours
verified	string	Verification status
time_and_compute_verification	string	Compute verification notes
link_to_colab_notebook	string	Google Colab notebook link
run_possible	string	Whether the code runs successfully
notes	string	Additional notes

Usage

from datasets import load_dataset

ds = load_dataset("AlgorithmicResearchGroup/aria-repo-benchmark", split="train")

for entry in ds:
    print(f"{entry['paper_title']} - {entry['modality']} - Reproducible: {entry['run_possible']}")

Related Resources

• ARIA Benchmark Suite
• Algorithmic Research Group - Open Source

Citation

@misc{aria_repo_benchmark,
    title={ARIA Repo Benchmark},
    author={Algorithmic Research Group},
    year={2024},
    publisher={Hugging Face},
    url={https://huggingface.co/datasets/AlgorithmicResearchGroup/aria-repo-benchmark}
}