datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
proof-bundlesNemotron-Math-Proofs-v3-SFT
Nemotron-Math-Proofs-v3-SFT
Dataset Description:
Nemotron-Math-Proofs-v3-SFT is a long-form mathematical reasoning dataset containing proof-generation, proof-refinement, verification, and meta-verification traces. The release contains 414,890 samples representing 15,818 unique problems after quality filtering.
The source pool contains 15,879 hard proof problems selected from the AoPS subset of nvidia/Nemotron-Math-Proofs-v1. Responses are generated using… See the full description on the dataset page: https://huggingface.co/datasets/nvidia/Nemotron-Math-Proofs-v3-SFT.cve-proof-corpus
CVE Proof Corpus
Six real vulnerability classes, each with a machine-checkable proof that the shipped fix
eliminates it — and a checker that shares no code with whatever produced the proof.
Every record carries the safety relation, the guard the upstream project shipped, the declared
attacker domain, and the nonnegative multipliers that prove the guard implies safety. All six verify.
pip install "certkit@git+https://github.com/nickharris808/certkit@main"
python verify.py… See the full description on the dataset page: https://huggingface.co/datasets/nickh007/cve-proof-corpus.Nemotron-Math-Proofs-v1
Nemotron-Math-Proofs-v1
Paper: Nemotron-Math: Efficient Long-Context Distillation of Mathematical Reasoning from Multi-Mode SupervisionCode: https://github.com/NVIDIA/NeMo-SkillsDocumentation: Nemotron-MathProofs-v1 documentation
Dataset Description:
Nemotron-Math-Proofs-v1 is a large-scale mathematical reasoning dataset containing ~580k natural language proof problems, ~550k formalizations into theorem statements in Lean 4, and ~900k model-generated reasoning… See the full description on the dataset page: https://huggingface.co/datasets/nvidia/Nemotron-Math-Proofs-v1.Nemotron-Math-Proofs-v2
Nemotron-Math-Proofs-v2
Dataset Description:
Nemotron-Math-Proofs-v2 is a mathematical proof-generation, verification, and meta-verification trace dataset. The problems are sourced from nvidia/Nemotron-Math-Proofs-v1 only taking the AoPS subset. The release contains 82,737 samples across 5,752 unique problems.
For this version, solutions are generated using DeepSeek-V4-Pro on Max inference mode. The generation pipeline produces proofs, verification traces, and… See the full description on the dataset page: https://huggingface.co/datasets/nvidia/Nemotron-Math-Proofs-v2.MA-ProofBench
MA-ProofBench: A Two-Tiered Evaluation of LLMs for Theorem Proving in Mathematical Analysis
English | 中文
We introduce MA-ProofBench, to the best of our knowledge, the first formal benchmark for evaluating large language models (LLMs) on theorem proving in Mathematical Analysis. It contains 200 rigorously formalized theorem-proving problems in Lean 4 + Mathlib (v4.28.0), split into two difficulty tiers:
Tier
Description
Source
Count
Level I
Undergraduate… See the full description on the dataset page: https://huggingface.co/datasets/openbmb/MA-ProofBench.lean-proofs-v1
Part of the SZL Holdings governed estate — claims are designed to carry checkable receipts. Verification proves integrity & origin, never accuracy or performance.
SZLHOLDINGS/lean-proofs-v1
The complete Lean 4 theorem library for the SZL Holdings Ouroboros Invariant research programme.
Doctrine v10/v11 Canonical Numbers
Metric
Value
Declarations
749
Unique axioms
14 (15 raw, 1 dup)
Sorries
163 (112 baseline + 51 Putnam)… See the full description on the dataset page: https://huggingface.co/datasets/SZLHOLDINGS/lean-proofs-v1.Nemotron-Math-Proofs-v3-RL
Nemotron-Math-Proofs-v3-RL
Dataset Description:
Nemotron-Math-Proofs-v3-RL is a long-form mathematical reasoning dataset for reinforcement learning. The release contains 9,597 proof-generation prompts.
The dataset uses NeMo Gym-compatible, single-turn user prompts derived from hard proof problems in the AoPS subset of nvidia/Nemotron-Math-Proofs-v1. The train split asks the policy to produce a rigorous solution and self-evaluation. Policy responses and realized… See the full description on the dataset page: https://huggingface.co/datasets/nvidia/Nemotron-Math-Proofs-v3-RL.proofjudge-eval-traces
ProofJudge Eval Traces
Agent interaction trajectories from ProofJudge evaluation runs,
stored in Harbor ATIF v1.6 format.
Each JSONL line is a complete trajectory of a judge agent evaluating one Lean 4 / Mathlib proof: the full
system and user prompts, every tool call it made to investigate the proof (search_mathlib, read_file,
pr_diff, …), every tool result it saw, and the structured verdict it returned.
Versions
This dataset has two generations, which must not… See the full description on the dataset page: https://huggingface.co/datasets/SJCaldwell/proofjudge-eval-traces.proofwriter-deduction-balancedA processed subset of the OWA section of the ProofWriter dataset.
Each train/test split contains 300 entries, each of which has a unique set of theories and a single question for those theories.
Both splits are balanced so that the depth of the proof required to answer the question varies evenly between 0-5 (50 entries each), and the labels are balanced (100 each).
'Unknown' labels have been replaced by 'Uncertain' to match other datasets.
proofwikips
ProofWiki Premise Selection Dataset, part of NaturalProofs Dataset
Sourced from NaturalProofs https://zenodo.org/records/4902289
ProofWriter
Github
https://github.com/teacherpeterpan/Logic-LLM/blob/main/outputs/logic_programs/ProofWriter_dev_gpt-4.json
Reference
@inproceedings{PanLogicLM23,
author = {Liangming Pan and
Alon Albalak and
Xinyi Wang and
William Yang Wang},
title = {{Logic-LM:} Empowering Large Language Models with Symbolic Solvers for Faithful Logical Reasoning},
booktitle = {Findings of the 2023 Conference on Empirical… See the full description on the dataset page: https://huggingface.co/datasets/renma/ProofWriter.sanctions-snapshot
ProofAML Sanctions & PEP Watchlist Snapshot
Research snapshot — not current data; production sanctions screening requires a live, monitored feed.
A resolved, deduplicated snapshot of global sanctions, PEP (politically exposed
persons), and watchlist entities, compiled by ProofAML
from government and intergovernmental publishers — OFAC, UN, EU, UK OFSI,
national parliaments, company registries, and more.
56 datasets, 118,770 entity records (version 2026-09-03)
Cross-source… See the full description on the dataset page: https://huggingface.co/datasets/proofaml/sanctions-snapshot.proofwriter-datasetgalahad-bench-proof
Galahad — SHA-256 Proof
Timestamped, independently verifiable measured benchmark results behind the Galahad paper.
📄 Paper: https://huggingface.co/papers/2607.23806
📚 arXiv: https://arxiv.org/abs/2607.23806
🧪 Live testbench (Space): https://huggingface.co/spaces/Corbenic/galahad-bench
💾 Proof mirror (GitHub): https://github.com/corbenicai/galahad-bench/tree/main/proof
What this is
This dataset is the integrity proof for the paper's claims: the raw measured… See the full description on the dataset page: https://huggingface.co/datasets/Corbenic/galahad-bench-proof.lean-proof-or-refute-300
Lean Proof-or-Refute 300
Lean Proof-or-Refute 300 is a compact collection of 300 formal reasoning
problems grounded in Lean 4 and Mathlib. Each problem starts from a verified
Mathlib theorem, makes one small numerical or operator mutation, and asks the
model to return either:
a Lean certificate proving the mutated proposition; or
a Lean certificate proving the exact negation of the complete proposition.
The model receives the related source theorem, a bounded source excerpt… See the full description on the dataset page: https://huggingface.co/datasets/xlr8harder/lean-proof-or-refute-300.proofnet-lean4lm-eval-results-Contamination-contaminated_proof_7b_v1.0_safetensor-private
Dataset Card for Evaluation run of Contamination/contaminated_proof_7b_v1.0_safetensor
Dataset automatically created during the evaluation run of model Contamination/contaminated_proof_7b_v1.0_safetensor
The dataset is composed of 62 configuration(s), each one corresponding to one of the evaluated task.
The dataset has been created from 4 run(s). Each run can be found as a specific split in each configuration, the split being named using the timestamp of the run.The "train"… See the full description on the dataset page: https://huggingface.co/datasets/nyu-dice-lab/lm-eval-results-Contamination-contaminated_proof_7b_v1.0_safetensor-private.ProofNet-Verified
ProofNet-Verified
Resource Links:
Paper: https://openreview.net/forum?id=5c0RSYyIWW, https://openreview.net/forum?id=NjgaeXNit3
GitHub: https://github.com/marcusm117/ProofNet-Verified
HuggingFace: https://huggingface.co/datasets/marcusm117/ProofNet-Verified
Visualization: https://marcusm117.github.io/ProofNet-Verified/
ProofNet-Verified (ProofNet-V) is an audited, corrected, and verified version of the ProofNet dataset that is upgraded to be compatible with Lean v4.28.0. A… See the full description on the dataset page: https://huggingface.co/datasets/marcusm117/ProofNet-Verified.douvras-lean-proof-repair
Douvras Lean Proof Repair Corpus
Exemplos sintéticos de erros comuns de reparo em Lean: importação ausente, incompatibilidade de
tipos, falha de tática, meta não resolvida, reescrita inválida e prova reflexiva. Os snippets não
foram executados no compilador (proof_status: NOT_EXECUTED); portanto o corpus não prova nenhum
teorema e não substitui validação com uma versão específica do Mathlib.
less-proofnet-lean4-top1Mproofjudge
ProofJudge
Paired Lean 4 / Mathlib human-authored proofs: the state a declaration was in when a pull request was
opened, and the state it was in when Mathlib merged it. A code-reviewing judge agent is aligned on a pair when it scores the merged version above the initial one.
Pairs were selected from Mathlib PRs by two filters: a mechanical one (at least 3 changed tokens and at least 5% change relative to the shorter proof) and an LLM judge (claude-sonnet-4-20250514) classifying… See the full description on the dataset page: https://huggingface.co/datasets/SJCaldwell/proofjudge.scbe-spine-overlay-proof-v1
Status: experimental. Experiment-specific slice. Primary public dataset: scbe-aethermoore-training-data.
SCBE Spine-Overlay Proof v1
Tiny demonstration bundle (18 rows = 3 domains x 6 tongues)
proving that the SCBE 12+ lane code-packet spine handles code,
chemistry, and mechanical motion as overlays on a single tokenizer
substrate, without forking the system.
Why this exists
Every row carries the same baseline lanes (binary, tokenizer, transport,
labels… See the full description on the dataset page: https://huggingface.co/datasets/issdandavis/scbe-spine-overlay-proof-v1.lm-eval-results-Contamination-contaminated_proof_7b_v1.0-private
Dataset Card for Evaluation run of Contamination/contaminated_proof_7b_v1.0
Dataset automatically created during the evaluation run of model Contamination/contaminated_proof_7b_v1.0
The dataset is composed of 62 configuration(s), each one corresponding to one of the evaluated task.
The dataset has been created from 4 run(s). Each run can be found as a specific split in each configuration, the split being named using the timestamp of the run.The "train" split is always pointing to… See the full description on the dataset page: https://huggingface.co/datasets/nyu-dice-lab/lm-eval-results-Contamination-contaminated_proof_7b_v1.0-private.ProofFlowBench
PROOFFLOWBENCH
A benchmark of 184 undergraduate-level mathematics theorems and proofs, each manually annotated with a step-by-step logical dependency graph (DAG), for evaluating proof autoformalization — the task of translating a natural-language proof into faithful, machine-verifiable Lean 4 code.
PROOFFLOWBENCH is the evaluation benchmark introduced in ProofFlow: A Dependency Graph Approach to Faithful Proof Autoformalization, accepted at ICLR 2026 (OpenReview · poster).… See the full description on the dataset page: https://huggingface.co/datasets/rafaelcabral96/ProofFlowBench.proofwikiqa
ProofWiki Question-Answer Retrieval Dataset
We use the theorems from the test set of the ProofWiki dataset in NaturalProofs as queries, and include all proofs from the dataset as the corpus.
amalia-Nemotron-Math-Proofs-v1
AMALIA Nemotron-Math-Proofs-v1
Version of the nvidia/Nemotron-Math-Proofs-v1 dataset used in the AMALIA's Supervised Fine-Tuning stage.
This dataset went through a processing pipeline to:
Remove entries with empty messages;
Drop extra unnecessary columns;
Original Dataset: https://huggingface.co/datasets/nvidia/Nemotron-Math-Proofs-v1
This dataset is provided as part of the AMALIA project and is included in the data mix used to post-train the AMALIA model.… See the full description on the dataset page: https://huggingface.co/datasets/amalia-llm/amalia-Nemotron-Math-Proofs-v1.Nemotron-Math-Proofs-v1
Dataset Description:
Nemotron-Math-Proofs-v1 is a large-scale mathematical reasoning dataset containing ~580k natural language proof problems, ~550k formalizations into theorem statements in Lean 4, and ~900k model-generated reasoning trajectories culminating in Lean 4 proofs. The dataset integrates human-authored problems with systematically generated formalizations and solution traces.
Each natural language problem is formalized by gpt-oss-120b into Lean 4 theorem statements with… See the full description on the dataset page: https://huggingface.co/datasets/Eduardo9991/Nemotron-Math-Proofs-v1.Nemotron-Math-Proofs-v1
Nemotron-Math-Proofs-v1
Paper: Nemotron-Math: Efficient Long-Context Distillation of Mathematical Reasoning from Multi-Mode SupervisionCode: https://github.com/NVIDIA/NeMo-SkillsDocumentation: Nemotron-MathProofs-v1 documentation
Dataset Description:
Nemotron-Math-Proofs-v1 is a large-scale mathematical reasoning dataset containing ~580k natural language proof problems, ~550k formalizations into theorem statements in Lean 4, and ~900k model-generated reasoning… See the full description on the dataset page: https://huggingface.co/datasets/ReactorJet/Nemotron-Math-Proofs-v1.Julia-Proof-Pile-2Julia-Proof-Pile-2
This dataset is part of Proof-Pile-2 dataset. This dataset is consisting of mathematical code, including numerical computing, computer algebra, and formal mathematics.
This entire dataset is in Julia language. It is slightly more than 0.5 Billion tokens. I have removed Meta data from this dataset hence you can directly use it for training purpose.
This dataset is in Jsonl format.