datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
hle_material_science
HLE Material Science: A Specialized Benchmark for Materials Science
A Materials Science Subset of Humanity's Last Exam (HLE)
Overview
HLE Material Science is a carefully curated materials science subset derived from the Humanity's Last Exam (HLE) dataset, containing 106 high-quality expert-level questions covering 25+ materials science subfields, with 97% of questions rated as high confidence.
This dataset is designed to evaluate large language models'… See the full description on the dataset page: https://huggingface.co/datasets/TalentZHOU/hle_material_science.natural-science-reasoning
Natural Sciences Reasoning: the "smolest" reasoning dataset
A smol-scale open dataset for reasoning tasks using Hugging Face Inference Endpoints. While intentionally limited in scale, this resource prioritizes:
Reproducible pipeline for reasoning tasks using a variety of models (Deepseek V3, Deepsek-R1, Llama70B-Instruct, etc.)
Knowledge sharing for domains other than Math and Code reasoning
In this repo, you can find:
The prompts and the pipeline (see the config file).
The… See the full description on the dataset page: https://huggingface.co/datasets/dvilasuero/natural-science-reasoning.llama-nemotron-science-reasoning-on-canonical-think-full
Llama-Nemotron science reasoning — Delphi canonical-think (COMPLETE, no length filter)
The complete reasoning:on science split of
nvidia/Llama-Nemotron-Post-Training-Dataset, converted once into the canonical
Delphi chat-template thinking format. 708,920 rows.
Unlike the cold-start warmup slice
open-athena/llama-nemotron-science-reasoning-on-le3000tok-100k
(and its -canonical-think variant), this build applies no length cap and no subsample — every
long-CoT science example is… See the full description on the dataset page: https://huggingface.co/datasets/laion/llama-nemotron-science-reasoning-on-canonical-think-full.Science-QnA
Science-QnA
The Science-QnA is a large-scale, high-quality science-focused dataset (~5.63M rows) curated using synthetic data generation through distillation techniques and select open-source resources. Designed to train and evaluate reasoning-capable models in science domains with emphasis on conceptual understanding, numerical problem-solving, and exam-style Q&A patterns across Physics, Chemistry, Biology, and Mathematics.
Summary
• Domain: Science, Physics… See the full description on the dataset page: https://huggingface.co/datasets/169Pi/Science-QnA.islamic-sciences
islamlab — The Islamic Sciences Corpus
The Islamic sciences other than Qur'an and hadith, as their authors wrote
them: 4,022 works by scholars who died between the
0st and the 14th Hijri century, cut along their own chapter
and biographical-entry boundaries into 1,864,389 units
(3.41 billion characters of Arabic), each carrying the volume and
page it sits on so a quotation can be cited rather than merely produced.
Scope is Ahl al-Sunnah wa'l-Jamāʿah, and the gate is the author… See the full description on the dataset page: https://huggingface.co/datasets/islamlab/islamic-sciences.openthoughts4-science-26041-prompts-qwen3-30b-a3B-thinking-2507-n8-flattened-logprobs-k16
OpenThoughts-4 Science SDG: Qwen3-30B-A3B-Thinking-2507 (n=8, top-16 logprobs)
Synthetic generations from
Qwen/Qwen3-30B-A3B-Thinking-2507
on the Marin OpenThoughts-4 science SDG prompt
set.
Each prompt is sampled n=8 times, and for every generated token the dataset
stores the chosen-token log probability plus the top-16 log probabilities
over the vocabulary, enabling distillation, KL-style fine-tuning,
reranking, and uncertainty analysis.
Generation setup
Field… See the full description on the dataset page: https://huggingface.co/datasets/marin-community/openthoughts4-science-26041-prompts-qwen3-30b-a3B-thinking-2507-n8-flattened-logprobs-k16.hle_material_science
HLE Material Science: A Specialized Benchmark for Materials Science
A Materials Science Subset of Humanity's Last Exam (HLE)
Overview
HLE Material Science is a carefully curated materials science subset derived from the Humanity's Last Exam (HLE) dataset, containing 106 high-quality expert-level questions covering 25+ materials science subfields, with 97% of questions rated as high confidence.
This dataset is designed to evaluate large language models'… See the full description on the dataset page: https://huggingface.co/datasets/stonelight/hle_material_science.openthoughts4-science-26041-prompts-qwen3-32b-n8-flattened-logprobs-k16
OpenThoughts-4 Science SDG: Qwen3-32B (n=8, top-16 logprobs)
Synthetic generations from
Qwen/Qwen3-32B
on the Marin OpenThoughts-4 science SDG prompt
set.
Each prompt is sampled n=8 times, and for every generated token the dataset
stores the chosen-token log probability plus the top-16 log probabilities
over the vocabulary, enabling distillation, KL-style fine-tuning,
reranking, and uncertainty analysis.
Generation setup
Field
Value
Generator model… See the full description on the dataset page: https://huggingface.co/datasets/marin-community/openthoughts4-science-26041-prompts-qwen3-32b-n8-flattened-logprobs-k16.task047_miscellaneous_answering_science_questions
Dataset Card for Natural Instructions (https://github.com/allenai/natural-instructions) Task: task047_miscellaneous_answering_science_questions
Additional Information
Citation Information
The following paper introduces the corpus in detail. If you use the corpus in published work, please cite it:
@misc{wang2022supernaturalinstructionsgeneralizationdeclarativeinstructions,
title={Super-NaturalInstructions: Generalization via Declarative Instructions on… See the full description on the dataset page: https://huggingface.co/datasets/Lots-of-LoRAs/task047_miscellaneous_answering_science_questions.Global-Ocean-Science-Corpus
🌊 Global-Ocean-Science-Corpus (v2.0 Curated & Cleaned)
A Highly Curated, Large-Scale Pre-Training & RAG Corpus for Deep Ocean Sciences, Marine Biology, and Oceanography
Language Note: This dataset is a 100% English-language scientific corpus (language: "en") aggregating peer-reviewed literature, deep-sea exploration dossiers, and technical oceanographic reports from leading global marine institutes.
Global-Ocean-Science-Corpus, derin okyanus bilimleri, deniz biyolojisi… See the full description on the dataset page: https://huggingface.co/datasets/tilikumotp/Global-Ocean-Science-Corpus.task701_mmmlu_answer_generation_high_school_computer_science
Dataset Card for Natural Instructions (https://github.com/allenai/natural-instructions) Task: task701_mmmlu_answer_generation_high_school_computer_science
Additional Information
Citation Information
The following paper introduces the corpus in detail. If you use the corpus in published work, please cite it:
@misc{wang2022supernaturalinstructionsgeneralizationdeclarativeinstructions,
title={Super-NaturalInstructions: Generalization via Declarative… See the full description on the dataset page: https://huggingface.co/datasets/Lots-of-LoRAs/task701_mmmlu_answer_generation_high_school_computer_science.data-science-en-id
Data Science EN-ID Parallel Corpus (Scientific Domain)
Dataset Description
This dataset is a curated English-Indonesian (EN-ID) parallel corpus specifically designed for the Scientific and Data Science domains. It was developed to support the training of Machine Translation (NMT) models and Large Language Models (LLMs) to better handle technical terminology, academic structures, and formal scientific language.
Primary Languages: English (EN) and Indonesian (ID)
Domain:… See the full description on the dataset page: https://huggingface.co/datasets/Ik45/data-science-en-id.tubitak-science-olympiad-tr
TUBITAK Science Olympiad Dataset
This dataset contains multiple-choice and open-ended scientific questions sourced from the TUBITAK (The Scientific and Technological Research Council of Turkey) Science Olympiads spanning various years. It is intended to serve as a benchmark for evaluating the advanced analytical, mathematical, and computational reasoning capabilities of Large Language Models (LLMs) in the Turkish language.
The dataset comprises approximately 2700 problems across… See the full description on the dataset page: https://huggingface.co/datasets/ytu-ce-cosmos/tubitak-science-olympiad-tr.islamic-sciences-training
islamlab — Islamic Sciences Training Sets
Training data derived from
islamlab/islamic-sciences:
text for domain adaptation, retrieval pairs with hard negatives, and citation
questions whose answers are read out of the corpus rather than written by a
model.
Nothing here is generated. Questions come from a fixed set of templates
and every answer is a field already present in the corpus. That buys a
narrow dataset in exchange for one that cannot teach a model a fact the
sources do… See the full description on the dataset page: https://huggingface.co/datasets/islamlab/islamic-sciences-training.scienceQA
Filter: no image && hint != ''
bvg_science_qwen_4b_not_easy
BVG Qwen3-4B science not-easy prompts
This is the exact two-split dataset artifact used by the BVG Qwen3-4B science
experiments. It is published as a DatasetDict with train and validation
splits.
Split
Rows
SHA-256 of canonical JSON rows
train
11,121
e985f63334809b43e8cffa971829bff6c2159fe08d79630b2fbbda9d22bc0831
validation
997
c2ea31a3e676b2c28c72daaddb9023c9163cd6671c7cf3afd2e305f7fc206480
The active experiment TOMLs consume the complete train split. Their… See the full description on the dataset page: https://huggingface.co/datasets/graf/bvg_science_qwen_4b_not_easy.openthoughts4-science-26041-prompts-qwen3-4b-n8-flattened-logprobs-k16
OpenThoughts-4 Science SDG: Qwen3-4B (n=8, top-16 logprobs)
Synthetic generations from
Qwen/Qwen3-4B
on the Marin OpenThoughts-4 science SDG prompt
set.
Each prompt is sampled n=8 times, and for every generated token the dataset
stores the chosen-token log probability plus the top-16 log probabilities
over the vocabulary, enabling distillation, KL-style fine-tuning,
reranking, and uncertainty analysis.
Generation setup
Field
Value
Generator model… See the full description on the dataset page: https://huggingface.co/datasets/marin-community/openthoughts4-science-26041-prompts-qwen3-4b-n8-flattened-logprobs-k16.science-cot-dataset
ExpertData Science — Scientific Reasoning
Expert-Annotated · Rights-Cleared · Ground-Truth Verified · PII-Clean
Each record captures a complete experimental or theoretical reasoning chain:
Hypothesis → Methodology → Causal Chain → Validated Conclusion.
Extracted from peer-reviewed papers across physics, biology, materials science, astrophysics, and neuroscience using structured scientific-reasoning extraction.
This dataset is produced by the ExpertData-Factory pipeline
(Mine →… See the full description on the dataset page: https://huggingface.co/datasets/expertdata-factory/science-cot-dataset.omni-science
GitHub
Website
Paper (Coming Soon)
Dataset Details
This dataset is a combination of corpora of text from scientific Wikipedia articles and scientific papers across major fields of science. This dataset contains continued-pretrain data.
Sources
This dataset was sourced from the following open-sourced datasets:
Science
zeroshot/arxiv-biology
legacy-datasets/wikipedia
bisectgroup/PubMed_TA
bluuebunny/biorxiv_abstract_embedding_mxbai_large_v1_milvus… See the full description on the dataset page: https://huggingface.co/datasets/omniomni/omni-science.k12-science-standards
[!WARNING]
Deprecated - use k12-science-standards-expanded instead.
This dataset is superseded: every instruction in this set also appears there, plus 1,123 more and nine additional metadata columns. Nothing here is unique to it.
It stays online so existing references keep resolving, but it will not be updated.
New work should point at robworks-software/k12-science-standards-expanded.
K-12 Science Standards (generated instruction data)
6,787 instruction/input/output records… See the full description on the dataset page: https://huggingface.co/datasets/robworks-software/k12-science-standards.wikipedia_science_chunked_small_rag_512
ScienceWikiSmallChunk
Processed version of millawell/wikipedia_field_of_science, prepared to be used in small context length RAG systems. Chunk length is tokenizer dependent, but each chunk should be around 512 tokens. Longer wikipedia pages have been split into smaller entries, with title added as a prefix.
There is also 256 tokens dataset available: Laz4rz/wikipedia_science_chunked_small_rag_256
If you wish to prepare some other chunk length:
use… See the full description on the dataset page: https://huggingface.co/datasets/Laz4rz/wikipedia_science_chunked_small_rag_512.IEEE2026_BigData_MAS-4-Science-Matching
SciAgentTrace
An execution-layer trace resource for scientific-agent workload characterization.
A protocol fixes who reasons, what each role can see, when feedback returns, and
when a workflow stops. Those choices determine the sequence of model requests
that produces an answer, so protocol design is also workload design. Two
workflows that consume similar token totals can issue very different request
sequences. SciAgentTrace records that difference.
The matched core runs the… See the full description on the dataset page: https://huggingface.co/datasets/AgentsSci/IEEE2026_BigData_MAS-4-Science-Matching.omni-science-instruct
GitHub
Website
Paper (Coming Soon)
Dataset Details
This dataset is a subset of a general STEM dataset, containing only science-related question-answer pairs. This dataset contains only chat-based data.
Sources
This dataset was sourced from the following open-sourced dataset:
Science
TIGER-Lab/WebInstructSub
wikipedia_dataset_science_en_id
Wikipedia Dataset Science (English - Indonesian)
Dataset Description
This dataset contains 122,433 aligned sentence pairs extracted from Wikipedia science articles in English and Indonesian. It is highly suitable for Natural Language Processing (NLP) tasks such as machine translation, cross-lingual alignment, and fine-tuning Large Language Models (LLMs) to better understand scientific terminology in Indonesian.
Language(s): English (en) and Indonesian (id)
Domain:… See the full description on the dataset page: https://huggingface.co/datasets/Ik45/wikipedia_dataset_science_en_id.synthetic-science-v2-sample
Synthetic Scientific Research Threads — v2 (sample)
A synthetic continual-learning benchmark: each episode is a coherent sequence of
short fictional scientific research documents about a single made-up entity, with
per-document QA anchors. Later documents build on, revise, or supersede earlier
ones. Designed to stress test-time / meta-learning approaches where a model must
adapt to a stream of documents and answer questions grounded in what it has just
seen.
This is a sample… See the full description on the dataset page: https://huggingface.co/datasets/HerrHruby/synthetic-science-v2-sample.wikipedia_science_chunked_small_rag_256
ScienceWikiSmallChunk
Processed version of millawell/wikipedia_field_of_science, prepared to be used in small context length RAG systems. Chunk length is tokenizer dependent, but each chunk should be around 256 tokens. Longer wikipedia pages have been split into smaller entries, with title added as a prefix.
There is also 512 tokens dataset available: Laz4rz/wikipedia_science_chunked_small_rag_512
If you wish to prepare some other chunk length:
use… See the full description on the dataset page: https://huggingface.co/datasets/Laz4rz/wikipedia_science_chunked_small_rag_256.synthlabs-GLM-5.2-Science
GLM-5.2 Science Synth Reasoning
Synthetic reasoning traces for science questions from the GLM-5.2 science dataset. Each record contains a complex scientific question with SYNTH-style reasoning and a generated answer.
Dataset Summary
33,014 records (605 dupes + 726 incomplete/truncated removed from 34,345 source)
33,014 reasoning turns (99.9% format compliance)
Average 3,094 chars per reasoning trace
Models Used
Model
Records… See the full description on the dataset page: https://huggingface.co/datasets/mkurman/synthlabs-GLM-5.2-Science.Mixture-Science
RecursiveMAS Mixture-Science
Project Page | Code | Paper
We introduce RecursiveMAS, a multi-agent framework that scales agent collaboration through latent-space recursion. This dataset contains training examples for the Mixture-Style setting.
Dataset Details
Item
Description
Dataset
RecursiveMAS/Mixture-Science
Original file
Mixture-Science.json
Collaboration style
Mixture-Style
Used for
science specialist inner agent training
Split
train
Rows… See the full description on the dataset page: https://huggingface.co/datasets/RecursiveMAS/Mixture-Science.OPV-Science
OPV Science: original experiment data
Prepared for Learning to Steer, Steering to See. Original four-domain split (physics, chemistry, biology, material). Original system/user prompts and answer-tag instructions are preserved.
Split
Rows
train
3,901
validation
434
Provenance and terms
The data is derived from the following sources; their original terms and attribution obligations continue to apply. No new blanket license is asserted over the… See the full description on the dataset page: https://huggingface.co/datasets/caiyuchen/OPV-Science.task688_mmmlu_answer_generation_college_computer_science
Dataset Card for Natural Instructions (https://github.com/allenai/natural-instructions) Task: task688_mmmlu_answer_generation_college_computer_science
Additional Information
Citation Information
The following paper introduces the corpus in detail. If you use the corpus in published work, please cite it:
@misc{wang2022supernaturalinstructionsgeneralizationdeclarativeinstructions,
title={Super-NaturalInstructions: Generalization via Declarative Instructions… See the full description on the dataset page: https://huggingface.co/datasets/Lots-of-LoRAs/task688_mmmlu_answer_generation_college_computer_science.