You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Research Use Only. This dataset contains model outputs and scores, not upstream benchmark content. Reproduce questions by running the released code against each upstream source under that source's own terms.

Log in or Sign Up to review the conditions and access this dataset content.

ARCK v0.1

Content-free model outputs (predictions/scores) for the ARCK v0.1 leaderboard — see IQuestLab/UBio-ARCK. Upstream questions are NOT included; regenerate them with the code in that repo and join by id (see Reproduction). Licensed CC-BY-NC-4.0, Research Use Only.

Record fields

field meaning
track top-level eval track: General / Co-Scientist / Bio-Designer
category ARCK sub-grouping within the track (e.g. Knowledge - Multi-domain)
subtask the specific eval task, e.g. mmlu_0shot_gen
leaf_id the leaderboard leaf node (benchmark unit; may bundle ≥1 subtask)
model model display name
tier publishing tier (see below): A / B / MINIMAL
source_ref upstream dataset name (joins the Sources table below); reproduce via that source
adaptation how upstream was adapted: none / prompt-adapted / task-reformulated
id per-item id (A/B: stable upstream-derived id; MINIMAL: opaque re-key)
prediction model's extracted answer/output (A/B only)
reasoning model's reasoning/CoT summary; null for General (A/B only)
gold correct upstream answer (A only)
score per-item score (float; null where no binary notion)
is_correct boolean correctness (MINIMAL only)

Tiers are our license-driven publishing classes, not upstream attributes:

  • A (permissive license): ships prediction, reasoning, gold
  • B (NonCommercial / ShareAlike / no-license): ships prediction, reasoning, no gold
  • MINIMAL (contamination-/redistribution-sensitive — GPQA/HLE/GISAID): only is_correct + score; no prediction/reasoning/gold, and id is an opaque re-key. (When loaded, those columns appear as null for MINIMAL rows — schema fill, not withheld content.)

Subsets and splits

Subset = model, split = track. Each of the 17 models is a config; within it the three tracks are splits (General, Co_Scientist, Bio_Designer — underscored to satisfy HF split naming).

load_dataset("IQuestLab/UBio-ARCK", "Kimi-K2.6")                  # all tracks for one model
load_dataset("IQuestLab/UBio-ARCK", "Kimi-K2.6", split="General") # one model, one track

Filter finer with the category / subtask fields.

Reproduction

Questions and prompts are regenerated, not shipped. To rebuild what each model was asked: download the upstream source (see the table below), run the conversion code in the UBio-ARCK repository, and join it onto these records by id. MINIMAL-tier items use opaque ids and are not rejoinable by design.

Sources, licenses, and reproduction prerequisites

Dataset License Source Reproduction prerequisite
AIME 2024 Apache-2.0 https://huggingface.co/datasets/HuggingFaceH4/aime_2024 Fetch from the linked upstream source.
AIME 2025 MIT https://huggingface.co/datasets/opencompass/AIME2025 Fetch from the linked upstream source.
ATLAS CC-BY-NC-SA-4.0 https://huggingface.co/datasets/opencompass/ATLAS Fetch from the linked upstream source.
BEACON Mixed https://drive.google.com/drive/folders/19ddrwI8ycvIxkgSV3gDo_VunLofYd4-6?usp=sharing Fetch from the linked upstream source.
ChemBench MIT https://huggingface.co/datasets/jablonkagroup/ChemBench Fetch from the linked upstream source.
DeepPrinciple MIT https://huggingface.co/deep-principle Fetch from the linked upstream source.
DROP CC-BY-SA-4.0 https://openaipublic.blob.core.windows.net/simple-evals/drop_v0_dev.jsonl.gz Fetch from the linked upstream source.
FrontierScience Apache-2.0 https://huggingface.co/datasets/openai/frontierscience Fetch from the linked upstream source.
GLRB CC-BY-NC-SA-4.0 https://huggingface.co/datasets/InstaDeepAI/genomics-long-range-benchmark Fetch from the linked upstream source.
GPQA CC-BY-4.0 https://huggingface.co/datasets/Idavidrein/gpqa Gated on HuggingFace — accept the authors' terms (incl. 'do not publish examples').
GPQADiamond CC-BY-4.0 https://openaipublic.blob.core.windows.net/simple-evals/gpqa_diamond.csv Fetch from the OpenAI simple-evals public mirror (no credentials).
GUE No license stated https://drive.google.com/file/d/1uOrwlf07qGQuruXqGXWMpPn8avBoW7T-/view?usp=sharing GUE leaf virus_covid requires your own GISAID access; other GUE leaves via upstream gdrive.
HLE MIT https://huggingface.co/datasets/cais/hle Gated on HuggingFace — accept the authors' terms; do not redistribute.
HumanEval MIT https://huggingface.co/datasets/openai/openai_humaneval Fetch from the linked upstream source.
IFEval Apache-2.0 https://huggingface.co/datasets/google/IFEval Fetch from the linked upstream source.
lab-bench CC-BY-SA-4.0 https://huggingface.co/datasets/futurehouse/lab-bench Fetch from the linked upstream source.
LiveCodeBench CC (unspecified) https://huggingface.co/datasets/livecodebench/code_generation_lite Fetch from the linked upstream source.
LiveProteinBench No license stated https://github.com/Rongdingyi/LiveProteinBench Fetch from the linked upstream source.
MATH 500 MIT https://huggingface.co/datasets/HuggingFaceH4/MATH-500 Fetch from the linked upstream source.
MMLU MIT https://huggingface.co/datasets/cais/mmlu Fetch from the linked upstream source.
MMLU-Pro MIT https://huggingface.co/datasets/TIGER-Lab/MMLU-Pro Fetch from the linked upstream source.
MoleculeNet MIT https://bioos-hermite-beijing.tos-cn-beijing.volces.com/unimol_data/finetune/molecular_property_prediction.tar.gz Fetch from the linked upstream source.
MoleculeQA MIT https://huggingface.co/datasets/IDEA-AI4S/MoleculeQA Fetch from the linked upstream source.
MolTextQA CC-BY-4.0 https://huggingface.co/datasets/sl160/MolTextQA Fetch from the linked upstream source.
mRNABench Mixed https://huggingface.co/collections/morrislab/mrnabench Fetch from the linked upstream source.
NT No license stated https://huggingface.co/datasets/InstaDeepAI/nucleotide_transformer_downstream_tasks_revised Fetch from the linked upstream source.
OligoGym CC-BY-4.0 https://huggingface.co/datasets/CollageBio/oligo-datasets Fetch from the linked upstream source.
PDFBench MIT https://github.com/pdfbench/PDFBench Fetch from the linked upstream source.
PRING MIT https://github.com/SophieSarceau/PRING Fetch from the linked upstream source.
ProteinGym MIT https://github.com/OATML-Markslab/ProteinGym Fetch from the linked upstream source.
RNAGym CC-BY-4.0 https://huggingface.co/datasets/Marks-lab/RNAgym Fetch from the linked upstream source.
s2bench No license stated https://huggingface.co/datasets/Duke-de-Artois/TOMG-Bench Fetch from the linked upstream source.
SciPredict MIT https://huggingface.co/datasets/ScaleAI/SciPredict Fetch from the linked upstream source.
SMolInstruct CC-BY-4.0 https://huggingface.co/datasets/osunlp/SMolInstruct Fetch from the linked upstream source.
T-Eval Apache-2.0 https://huggingface.co/datasets/lovesnowbest/T-Eval Fetch from the linked upstream source.
TAPE BSD-3-Clause https://github.com/songlab-cal/tape Fetch from the linked upstream source.
VenusX Apache-2.0 https://huggingface.co/collections/AI4Protein/venusx-dataset Fetch from the linked upstream source.

Citation

@misc{arck2026,
  title  = {ARCK: The Arc of Scientific Intelligence},
  author = {IQuest Research UBio Team},
  year   = {2026},
}
Downloads last month
-