ScienceClaw-Eval

The evaluation data of ScienceClaw-Eval, the companion benchmark of ScienceClaw, an agent system that improves from replay-verified executions by evolving its Skills and typed Operators. Project page: scienceclaw.science.

ScienceClaw-Eval measures continual self-evolution of AI-for-Science agents rather than single-shot ability. It covers 23 disciplines across the natural and social sciences (ANZSRC Fields of Research, FoR30–FoR52). For every discipline, a task-native metric and a set of scientific hard constraints decide whether an instance counts as solved.

What is in this dataset

This dataset contains 46 flat source slots: FoR30–FoR52, each with iid and ood, and exactly 64 records per slot (2,944 records total).

Split	Meaning
iid	In-distribution held-out instances.
ood	Instances from an independently sourced dataset of the same discipline, used to test cross-dataset transfer.

Each slot is under forXX/iid/ or forXX/ood/ and has metadata.jsonl. Each row is labelled with source_dataset, sample_id, source_url, license, split, and metric, plus the paths of its payload files. Compact payloads are under that slot's data/ directory. There are no agent episodes, traces, or model outputs. A few slots also ship the support files their official evaluation needs: the 24 source-normal clips of FoR40 OOD, and the source/validation inputs and targets under for42/iid/data/evaluator/.

Some sources publish binary samples under their original terms; those rows retain the stable source ID and cited URL so the original data can be obtained from the authoritative source.

The metric field follows the paper table for the native routes; FoR30 OOD uses the related PQ_leaf metric while FoR40 keeps DCASE: PQ+, Spearman, DSC, NRMSE (%), ROC-AUC, MASE, SDR (dB), RMSE (K), sMAPE (%), 10-mask acc., DCASE, Clin. utility, cMER-micro, nRMSE, chrF++, pass@1, LAS, mAP, Oracle acc., F1, MAE, and Micro acc.

The 23 disciplines

Slot	Discipline	Task	Metric	IID source	OOD source
for30	Agricultural, veterinary and food sciences	plant and leaf panoptic segmentation	PQ+ ↑	PhenoBench v1.1.0 official	CropAndWeedAndLeaf
for31	Biological sciences	protein variant fitness ranking	Spearman ↑	ProteinGym (substitution assays)	TAPE fluorescence (Sarkisyan et al. 2016)
for32	Biomedical and clinical sciences	hippocampus segmentation in MRI	DSC ↑	MSD Task04 Hippocampus	UCL Clark-Maguire hippocampal subfield segmentation (Dryad)
for33	Built environment and design	building load forecasting	NRMSE (%) ↓	BuildingsBench	End-Use Load Profiles (EULP)
for34	Chemical sciences	molecular activity classification	ROC-AUC ↑	OGB ogbg-molhiv	BACE (MoleculeNet)
for35	Commerce, management, tourism and services	tourism series forecasting	MASE ↓	Monash Tourism Monthly	Monash Tourism Quarterly
for36	Creative arts and writing	music source separation	SDR (dB) ↑	MUSDB18	MoisesDB
for37	Earth sciences	2 m temperature forecasting	RMSE (K) ↓	WeatherBench 2 (ERA5 2 m temperature, 2019)	WeatherBench 2 (ERA5 2 m temperature, 2020)
for38	Economics	macroeconomic forecasting	sMAPE (%) ↓	World Bank WDI	World Bank WDI
for39	Education	adaptive educational testing	10-mask accuracy ↑	Eedi NeurIPS 2020 Education Challenge (Task 4)	EdNet-KT1
for40	Engineering	anomalous sound detection	DCASE score ↑	DCASE 2024 Task 2 official evaluation (Zenodo 11363076)	DCASE 2023 Task 2 ToyNscale evaluation
for41	Environmental sciences	probabilistic aquatic forecasting	CRPS ↓	NEON aquatics (neon4cast)	USGS river metabolism
for42	Health sciences	sepsis early warning	clinical utility ↑	PhysioNet/CinC Challenge 2019 (Sepsis)	SepsisExp: Patient Timelines with Expert Sepsis Labels
for43	History, heritage and archaeology	OCR post-correction	cMER-micro ↓	HIPE-OCRepair-2026 (icdar2017, impresso-nzz, impresso-snippets)	ICDAR 2019 POCR
for44	Human society	causal treatment-effect estimation	nRMSE ↓	ACIC 2016	IHDP
for45	Indigenous studies	Indigenous-language captioning	chrF++ ↑	AmericasNLP 2026	Bloom Captioning (Mam)
for46	Information and computing sciences	code generation	pass@1 ↑	HumanEval	SWE-bench Verified
for47	Language, communication and culture	dependency parsing	LAS ↑	Universal Dependencies (Marathi-UFAL, Italian-ParTUT, Arabic-PADT, Croatian-SET; 16 sentences each)	Universal Dependencies (Indonesian-GSD, Swedish_Sign_Language-SSLC, French-GSD, Hindi-HDTB; 16 sentences each)
for48	Law and legal studies	contract evidence retrieval	mAP ↑	ContractNLI	ACORD (Atticus Clause Retrieval Dataset)
for49	Mathematical sciences	SMT satisfiability prediction	oracle-agreement accuracy ↑	SMT-LIB 2025 (QF_NIA)	SMT-LIB 2024 (QF_NIA)
for50	Philosophy and religious studies	human-value detection	F1 ↑	Touché23-ValueEval	ETHICS (commonsense)
for51	Physical sciences	phonon property prediction	MAE ↓	Matbench phonons	Kyoto PhononDB
for52	Psychology	human choice prediction	micro accuracy ↑	Psych-201	Psych-101

Loading

Fetch one discipline, or one split of it:

import json
from huggingface_hub import snapshot_download

root = snapshot_download(
    "beita6969/scienceclaw-eval",
    repo_type="dataset",
    allow_patterns=["for35/ood/*"],        # Commerce and services, OOD
)
rows = [json.loads(l) for l in open(f"{root}/for35/ood/metadata.jsonl")]
print(len(rows), rows[0]["source_dataset"], rows[0]["metric"], rows[0]["files"])

Using it with ScienceClaw

The agent system is in the ScienceClaw repository; it ships a discipline skill for each of these slots (scienceclaw-benchmark-for30 … for52) that describes the task family: inputs, deliverable, how quality is judged, and the tools that fit. This dataset holds the data only; the repository contains no evaluation data.

Licenses and attribution

Every row carries the license and source_url of its original source, and the original terms apply. Some sources are non-commercial or share-alike (for example, FoR30 OOD and FoR40 are CC BY-NC-SA 4.0). Check the row-level license before reuse, and obtain restricted data from the authoritative source where noted.

Contact

mingdazhang@ieee.org