datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
ResearchMath-14k
ResearchMath-14k
ResearchMath-14k is a collection of 14,056 research-level mathematical problem records extracted from papers, open-problem lists, workshop sheets, and related academic sources. Each record contains the original extracted question, a rewritten self-contained problem statement, taxonomy labels, and open-status metadata.
Paper: ResearchMath-14K: Scaling Research-Level Mathematics via Agents
Load
from datasets import load_dataset
ds =… See the full description on the dataset page: https://huggingface.co/datasets/amphora/ResearchMath-14k.ResearchMath-Reasoning-194K
ResearchMath-Reasoning-194K
ResearchMath-Reasoning-194K is a collection of 193,938 long-form reasoning traces and solutions for research-level mathematical problems, released alongside ResearchMath-14k as part of the same paper. While ResearchMath-14k provides the curated problem statements, this dataset provides model-generated solution attempts: each record contains a self-contained problem statement, a long chain-of-thought reasoning trace, and a final response.
Paper:… See the full description on the dataset page: https://huggingface.co/datasets/amphora/ResearchMath-Reasoning-194K.QwQ-LongCoT-130KAlso have a look on the second version here => QwQ-LongCoT-2
Figure 1: Just a cute picture generate with [Flux](https://huggingface.co/Shakker-Labs/FLUX.1-dev-LoRA-Logo-Design)
Today, I’m excited to release QwQ-LongCoT-130K, a SFT dataset designed for training O1-like large language models (LLMs). This dataset includes about 130k instances, each with responses generated using QwQ-32B-Preview. The dataset is available under the Apache 2.0 license, so feel free to use it as you like.… See the full description on the dataset page: https://huggingface.co/datasets/amphora/QwQ-LongCoT-130K.dasd-stage1-50k
DASD stage1 - 50k length-filtered subset
A 50,000-example subset of the stage1 (low-temperature) config of
Alibaba-Apsara/Superior-Reasoning-SFT-gpt-oss-120b.
Columns are input / output; output is the verbatim gpt-oss-120b <think> reasoning trace.
How it was built
Started from stage1 (104,829 rows).
Applied the Qwen3-4B-Instruct-2507 chat template and tokenized the full formatted
conversation, then dropped every example over 65,536 tokens (the 64K training… See the full description on the dataset page: https://huggingface.co/datasets/amphora/dasd-stage1-50k.ample-math
AMPLE-Math
5,319 mathematics problems, each with a verified final answer and six references to that same
answer. The references differ only in how much of the reasoning they show, which makes them useful
for studying what a teacher's reference content contributes during distillation.
Problems and original reasoning come from the metadata configuration of
OpenThoughts-114k, and keep its
Apache-2.0 attribution. A question was kept only if all six references exist, every generated… See the full description on the dataset page: https://huggingface.co/datasets/xiuyuz/ample-math.nemotron-math-v4-dedup-86k
Nemotron-SFT-Math-v4 - deduplicated, length-filtered
Filtered subset of nvidia/Nemotron-SFT-Math-v4,
as {input, output} with output = <think>{reasoning}</think>{answer}.
step
rows
source
545,431
drop non-empty tools
285,516
drop assistant > 65,536 tok
272,576
keep the shortest solution per problem
86,177
drop full conversation > 65,536 tok
86,161
Tokenizer Qwen/Qwen3-4B-Instruct-2507 (shared with Qwen3-30B-A3B-Instruct-2507). The source
reasoning_content… See the full description on the dataset page: https://huggingface.co/datasets/amphora/nemotron-math-v4-dedup-86k.ResearchMath-Filtered
ResearchMath-Filtered
ResearchMath-Filtered is a quality-filtered collection of 129,927 long-form reasoning traces and
solutions for research-level mathematical problems, released alongside
ResearchMath-14k as part of the same
paper. It is a cleaned subset of
ResearchMath-Reasoning-194K:
each record holds a self-contained problem statement, a long chain-of-thought reasoning trace, and a
final response, with low-quality and non-solving generations removed.
Paper:… See the full description on the dataset page: https://huggingface.co/datasets/amphora/ResearchMath-Filtered.math-intuition-reasoning-traces
math-intuition reasoning traces
Full chain-of-thought traces from 7 reasoning models on the same 4,020 problems, graded
by each problem family's own verifier.
Questions come from
amphora/math-intuition-20260908-402-easy-10
— 402 arXiv-derived problem families x 10 seeds, easy preset. Every row here refers to an id
in that dataset, so prompts and the instance cache can be joined from it.
Generation settings
Identical for every model, so the traces are directly… See the full description on the dataset page: https://huggingface.co/datasets/amphora/math-intuition-reasoning-traces.ArXivOpenProblems
ArXivOpenProblems
294,444 self-contained research questions mined from the future-work, open-problem
and limitation statements of 82,376 arXiv mathematics papers.
Each row pairs a verbatim quote from a paper with a standalone research question
rewritten so that it can be read and understood without the source paper in hand.
Fields
field
description
uid
<arxiv_id>#<index> — identifier of the finding within its paper
arxiv_id
arXiv identifier… See the full description on the dataset page: https://huggingface.co/datasets/amphora/ArXivOpenProblems.ResearchMath-77k
ResearchMath-77k (Qwen3-30B-A3B boxed slice)
Clean SFT training slice used for amphora/llama-rm-trained and amphora/qwen3-4b-boxed-after-think-77k.
A filtered subset of amphora/ResearchMath-Reasoning-194K:
model == "Qwen/Qwen3-30B-A3B" (100,512 rows)
final response contains a \boxed{...} answer (~77,200 rows)
formatted think/response sequence under 32,768 Llama-3.1 tokens (~77,193 rows)
Columns
problem research-level math problem statement
reasoning long… See the full description on the dataset page: https://huggingface.co/datasets/amphora/ResearchMath-77k.LessWrong-Amplify-Instruct
This is the Official LessWrong-Amplify-Instruct dataset. Over 500 multi-turn examples, and many more coming soon!
This leverages Amplify-Instruct method to extend thousands of scraped Less-Wrong posts into advanced in-depth multi-turn conversations.
Comprised of over 500 highly filtered multi-turn synthetic conversations.
Average context length per conversation is over 2,000 tokens. (will measure this more accurately soon)
Synthetically created using a newly developed pipeline… See the full description on the dataset page: https://huggingface.co/datasets/LDJnr/LessWrong-Amplify-Instruct.math-intuition-20260906-403-demo-10
math-intuition-20260906-403-demo-10
3,936 mathematics problems drawn from 403 problem families, each derived from a
distinct arXiv paper. Every problem is generated answer-first, so the answer is known by
construction and is checked by the family's own verify() before the row is written.
No row in this file is ungraded.
This is the demo rung — read this before using it
Each family exposes a four-rung ladder: demo, easy, medium, hard. This file samples
demo, which… See the full description on the dataset page: https://huggingface.co/datasets/amphora/math-intuition-20260906-403-demo-10.math-intuition-20260906-403-easy-30
math-intuition-20260906-403-easy-30
12,090 synthetic mathematics problems drawn from 403 problem families, each family
derived from a distinct arXiv paper. Every problem is generated answer-first, so the
answer is known by construction and is checked by the family's own verify() before
the row is written. No row in this file is ungraded.
This is the easy slice: 30 instances per family at each family's easiest difficulty
preset. It is not the hard benchmark — see Difficulty… See the full description on the dataset page: https://huggingface.co/datasets/amphora/math-intuition-20260906-403-easy-30.