EEGAgentBenchmark

EEGAgentBenchmark is the frozen data release for EEGAgentBench, a benchmark of general-purpose LLM agents on EEG knowledge and signal-analysis tasks. The benchmark contains 1,072 evaluation instances across six tasks. This repository provides the 1,034 EEG/PSG signal files used by T2--T6; T1 has no signal file.

The benchmark is designed for tool-using agents. Signal files and answer keys are resolved by the runtime and are not inserted into model prompts. Labels are public for reproducible local evaluation, so this release is not intended to serve as a hidden-label leaderboard.

Tasks

Task	Description	Source	Instances	Signal files	Primary metric
T1	EEG Knowledge QA	MedMCQA-EEG-strict	38	0	Accuracy
T2	Artifact Identification	EEGdenoiseNet	300	300	Macro-F1
T3	Normal-vs-Epileptic Screening	Bonn	188	188	Macro-F1
T4	Dementia Cohort Classification	OpenNeuro ds004504	69	69	Macro-F1
T5	Seizure Event Detection	CHB-MIT	280	280	Event-F1
T6	Sleep Staging	Sleep-EDFx	197	197	Macro-F1

The 1,034 signal files total 24,244,274,188 bytes (approximately 23 GiB).

Repository Contents

data/
├── eegdenoisenet/       # 300 NumPy signal arrays for T2
├── bonn/                # 188 text signals for T3
├── ds004504/            # 69 EEGLAB .set recordings for T4
├── chbmit/              # 280 EDF recordings for T5
└── sleep_edfx/          # 197 frozen EDF recordings for T6
metadata/
├── files.jsonl          # path, byte size, and SHA-256 for every signal
├── instances.jsonl      # all 1,072 instance-to-signal mappings
├── summary.json         # task-level counts and manifest hashes
└── t6_trim_plan.jsonl   # deterministic T6 reconstruction/cropping plan

t6_trim_plan.jsonl is provenance metadata. It is not needed when using the already frozen T6 EDF files, but records how to reproduce them from the upstream Sleep-EDFx source files. Of the 197 T6 records, 155 are cropped and 42 are retained without cropping.

Download and Use

Install the Hugging Face CLI, then download data/ and metadata/ into the root of the EEGAgentBench code release:

pip install -U huggingface_hub
hf download whhhy123/EEGAgentBenchmark \
  data/ metadata/ \
  --repo-type dataset \
  --local-dir /path/to/EEGAgentBench
cd /path/to/EEGAgentBench
eeg-bench prepare --fast

For a complete content-hash verification, run:

python scripts/validate_release.py --hash

The companion code release contains the public scenarios, answer keys, runtime manifests, deterministic signal tools, official metrics, and model adapters. This data repository intentionally contains only the large signal files and their provenance metadata; downloading it alone does not provide the runner or metric implementation.

Data Sources and Terms

This is a mixed-source redistribution. The repository-level license: other field does not replace the terms of the upstream datasets. Each file remains subject to its source dataset's license or terms, and users must retain the required attribution.

Benchmark data	Upstream source	Version	Terms recorded by this release
T2	EEGdenoiseNet	1.0	CC0-1.0
T3	Bonn EEG dataset (R. G. Andrzejak et al.)	1.0	Research and education only; no commercial or military use
T4	OpenNeuro ds004504	1.0.8	CC0-1.0
T5	CHB-MIT	1.0.0	ODC-By-1.0
T6	Sleep-EDFx	1.0.0	ODC-By-1.0

See LICENSES/README.md for attribution and license links. Users should also cite the EEGAgentBench paper and the relevant upstream datasets in derived work.

Integrity and Reproducibility

• metadata/files.jsonl is the authoritative list of all signal files.
• Each entry includes the exact byte size and SHA-256 digest.
• metadata/instances.jsonl maps every evaluation instance to its release path.
• Public model prompts do not contain answer labels or local filesystem paths.
• Answer keys in the companion code release are used only by the evaluator and must never be included in a model prompt.
• Official scores must be computed with the companion evaluator rather than by reimplementing aggregation from model outputs.

Intended Use and Limitations

The release is intended for reproducible research on EEG-oriented tool-using agents. It is not a clinical device and must not be used for diagnosis, patient care, or treatment decisions. Results depend on the supplied tools, prompts, model configuration, API behavior, and runtime environment. Because labels are public, users are responsible for preventing answer-key leakage into prompts or model-development data when reporting benchmark results.