Error code:   ConfigNamesError
Exception:    ComposerError
Message:      expected a single document in the stream
  in "<unicode string>", line 1, column 1:
    pretty_name: AIFaultBench
    ^
but found another document
  in "<unicode string>", line 29, column 1:
    --- 
    ^
Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/job_runners/dataset/config_names.py", line 67, in compute_config_names_response
                  config_names = get_dataset_config_names(
                      path=dataset,
                      token=hf_token,
                  )
                File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 161, in get_dataset_config_names
                  dataset_module = dataset_module_factory(
                      path,
                  ...<4 lines>...
                      **download_kwargs,
                  )
                File "/usr/local/lib/python3.14/site-packages/datasets/load.py", line 1217, in dataset_module_factory
                  raise e1 from None
                File "/usr/local/lib/python3.14/site-packages/datasets/load.py", line 1192, in dataset_module_factory
                  ).get_module()
                    ~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/load.py", line 601, in get_module
                  dataset_card_data = DatasetCard.load(dataset_readme_path).data
                                      ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/huggingface_hub/repocard.py", line 187, in load
                  return cls(f.read(), ignore_metadata_errors=ignore_metadata_errors)
                File "/usr/local/lib/python3.14/site-packages/huggingface_hub/repocard.py", line 77, in __init__
                  self.content = content
                  ^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/huggingface_hub/repocard.py", line 95, in content
                  data_dict = yaml.safe_load(yaml_block)
                File "/usr/local/lib/python3.14/site-packages/yaml/__init__.py", line 125, in safe_load
                  return load(stream, SafeLoader)
                File "/usr/local/lib/python3.14/site-packages/yaml/__init__.py", line 81, in load
                  return loader.get_single_data()
                         ~~~~~~~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/yaml/constructor.py", line 49, in get_single_data
                  node = self.get_single_node()
                File "/usr/local/lib/python3.14/site-packages/yaml/composer.py", line 41, in get_single_node
                  raise ComposerError("expected a single document in the stream",
                          document.start_mark, "but found another document",
                          event.start_mark)
              yaml.composer.ComposerError: expected a single document in the stream
                in "<unicode string>", line 1, column 1:
                  pretty_name: AIFaultBench
                  ^
              but found another document
                in "<unicode string>", line 29, column 1:
                  --- 
                  ^

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

AIFaultBench: A Benchmark of Real-World Faults in AI Software Systems

AIFaultBench is a benchmark of 770 real-world AI software faults collected from 105 open-source repositories across 76 organizations, spanning traditional machine learning, deep learning, large language model infrastructure, reinforcement learning, agentic AI systems, and AI tooling.

Each fault includes everything needed for reproduction:

GitHub issue report
Minimal reproduction script
Dependency specification
Codebase reconstruction script
Environment setup script
Reproduction logs
Structured metadata
Reproduction trajectory

652 (85%) faults are verified reproducible, while the remaining faults include documented reasons preventing reproduction.

Bugs	Reproducible	Repositories	Organizations	Domains
770	652	105	76	6

Dataset Overview

Domain	Bugs	Reproducible
Deep Learning	307	83%
LLM Infrastructure	151	75%
Agentic AI	130	92%
Machine Learning	118	86%
Reinforcement Learning	40	92%
AI Tooling	24	100%

Repository Structure

.
├── index.json
├── index.csv
├── consume_dataset.ipynb
├── bugs/
│   └── <bug_id>/
│       ├── bug_report.txt
│       ├── repro.py
│       ├── requirements.txt
│       ├── setup_codebase.sh
│       ├── setup_env.sh
│       ├── run_repro.sh
│       ├── reproduction.json
│       ├── reproduction_trajectory.md
│       └── logs/

The benchmark is indexed through both index.json and index.csv, while each bug is packaged independently inside bugs/<bug_id>.

Quick Start

Reproducing a fault requires only Git, Python, and a POSIX-compatible shell.

cd bugs/<bug_id>
bash setup_codebase.sh
bash run_repro.sh

The scripts automatically

reconstruct the buggy codebase,
create an isolated Python environment,
install dependencies,
execute the reproduction script.

No Docker or API credentials are required.

Loading the Dataset

import json

bugs = json.load(open("index.json"))

reproducible = [b for b in bugs if b["reproducible"]]
agentic = [b for b in bugs if b["domain"] == "Agentic"]

The index is also exposed through the Hugging Face dataset viewer:

from datasets import load_dataset

index = load_dataset("mehilshah/AIFaultBench", split="train")

Note that bug_id is a zero-padded three-digit string (001, …, 774) that names the directory under bugs/. The CSV loader parses it as an integer, so pad it back before building a path:

path = f"bugs/{int(row['bug_id']):03d}"

Loading index.json directly preserves the identifier as a string.

For a complete walkthrough, see consume_dataset.ipynb.

Package Contents

Each bug contains

original GitHub issue
minimal reproduction script
dependency specification
environment setup
codebase reconstruction
execution logs
structured reproduction metadata
reproduction trajectory

Together, these provide a fully executable reproduction package suitable for evaluating debugging, fault localization, automated repair, bug reproduction, and AI software engineering tools.

Citation

@misc{AIFaultBench_2026,
  title={AIFaultBench: A Reproducible Benchmark of Real-World AI Software Faults},
  author={Shah, Mehil B and Rahman, Mohammad Masudur and Khomh, Foutse},
  year={2026},
  doi={10.5281/zenodo.21782307},
  url={https://zenodo.org/records/21782307}
}

License

The reproduction scripts, metadata, and benchmark packaging are released under CC BY 4.0.

Each original bug report and source repository remains under its respective upstream license.

Downloads last month: 883

Total file size:

31.2 MB