datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
sat-vl-sft-training-ready-v1
Dataset Summary
NuTonic/sat-bbox-metadata-sft-v1 is a metadata-first, procedural VLM SFT dataset built from an existing “sat-bbox” style dataset tree (Sentinel‑2 chips + per-tile JSON metadata sidecars, optionally paired Mapbox stills).
The goal is to create high-signal, production-shaped supervision for multimodal chat models:
Captioning for satellite chips
Grounding (bounding boxes in normalized coordinates) for land-cover regions
Class-focused captions and absence checks for… See the full description on the dataset page: https://huggingface.co/datasets/NuTonic/sat-vl-sft-training-ready-v1.sat-vl-sft-postprocessed-merged-v1
Dataset Summary
NuTonic/sat-bbox-metadata-sft-v1 is a metadata-first, procedural VLM SFT dataset built from an existing “sat-bbox” style dataset tree (Sentinel‑2 chips + per-tile JSON metadata sidecars, optionally paired Mapbox stills).
The goal is to create high-signal, production-shaped supervision for multimodal chat models:
Captioning for satellite chips
Grounding (bounding boxes in normalized coordinates) for land-cover regions
Class-focused captions and absence checks for… See the full description on the dataset page: https://huggingface.co/datasets/NuTonic/sat-vl-sft-postprocessed-merged-v1.NuminaMath-LEAN-satp-v4.27
NuminaMath-LEAN-satp-v4.27
Lean 4 formal-statement + initial proof goal_state pairs over the
NuminaMath-LEAN problem pool, packaged for Lean 4.27.0. This is the main
training set for SATP (Steering Aesop for Theorem Proving) running in a
Lean 4.27 environment.
Every row's formal_statement elaborates cleanly under the pinned toolchain
below, and every goal_state is the pretty-printed goal produced in that
row's own environment — the same rendering the SATP runtime observes at… See the full description on the dataset page: https://huggingface.co/datasets/ChristianZ97/NuminaMath-LEAN-satp-v4.27.NuminaMath-LEAN-satp-buffer-dspaug-Temp
NuminaMath-LEAN-satp-buffer-dspaug-Temp
Staging buffer for the DSP+ paper-augmentation sweep (2026-04-29). This
is a -Temp variant — lemma_names / lemma_scores are empty and
theorem_uuid is the join key (matches NuminaMath-LEAN-satp.uuid).
Retrieval population + rename-to-uuid happens at the
promote-to-canonical merge step, mirroring the precedent set by
NuminaMath-LEAN-satp-buffer-planf-v1-Temp.
Why this exists
Schema audit on NuminaMath-LEAN-satp-buffer (40,965… See the full description on the dataset page: https://huggingface.co/datasets/ChristianZ97/NuminaMath-LEAN-satp-buffer-dspaug-Temp.sat-bbox-metadata-sft-v1
Dataset Summary
NuTonic/sat-bbox-metadata-sft-v1 is a metadata-first, procedural VLM SFT dataset built from an existing “sat-bbox” style dataset tree (Sentinel‑2 chips + per-tile JSON metadata sidecars, optionally paired Mapbox stills).
The goal is to create high-signal, production-shaped supervision for multimodal chat models:
Captioning for satellite chips
Grounding (bounding boxes in normalized coordinates) for land-cover regions
Class-focused captions and absence checks for… See the full description on the dataset page: https://huggingface.co/datasets/NuTonic/sat-bbox-metadata-sft-v1.Muse-Glimmer-SWE-Gym-2k
Muse-Glimmer-SWE-Gym-2k
Agentic coding traces from meta-models/Muse-Glimmer-30B, recorded for training a
speculative-decoding drafter. 1,981 mini-swe-agent trajectories over SWE-Gym and
SWE-bench-extra instances, and the 159,999 individual chat-completion calls behind them.
Configs
Config
Rows
Size
What it is
train
1,981
57 MB
One row per trajectory: the full conversation as messages.
raw
159,999
2.7 GB
One row per recorded API call: request and… See the full description on the dataset page: https://huggingface.co/datasets/Satgoy152/Muse-Glimmer-SWE-Gym-2k.math_benchmark_test_saturation
LLM Leaderboard Data for Hendrycks MATH Dataset (2022–2024)
This dataset aggregates yearly performance (2022–2024) of large language models (LLMs) on the Hendrycks MATH benchmark. It is specifically compiled to explore performance evolution, benchmark saturation, parameter scaling trends, and evaluation metrics of foundation models solving complex math word problems.
Original source data: Math Word Problem Solving on MATH (Papers with Code)
About Hendrycks' MATH… See the full description on the dataset page: https://huggingface.co/datasets/nlile/math_benchmark_test_saturation.minif2f-satp-alphaproof
minif2f-satp-alphaproof
The canonical 488-base-problem view of the miniF2F benchmark used by
AlphaProof, ported to Lean 4.26.0 and packaged with initial proof
goal_state values generated for SATP v2.
This dataset is intended as the held-out evaluation (test) and
hyperparameter-tuning (validation) benchmark for SATP / sketch-and-prove
pipelines running in the same Lean 4.26 environment.
The source is Google DeepMind
miniF2F@f0a20e1.
Its README identifies this as the benchmark… See the full description on the dataset page: https://huggingface.co/datasets/ChristianZ97/minif2f-satp-alphaproof.NuminaMath-LEAN-satp
NuminaMath-LEAN-satp
Lean 4 formal-statement + initial proof goal_state pairs harvested from
the NuminaMath-LEAN problem pool. This is the main training set for
SATP (Steering Aesop for Theorem Proving) and the target distribution
that all sibling datasets in this collection align with byte-for-byte.
Sibling datasets (same uuid scheme so they join cleanly):
NuminaMath-LEAN-satp-gaps — augmented train set with sub-goal (gap) records harvested from verified sketches… See the full description on the dataset page: https://huggingface.co/datasets/ChristianZ97/NuminaMath-LEAN-satp.Qwen3.5-0.8B-selfdistill
Qwen3.5-0.8B self-distillation pairs (DSpark drafter training data)
The exact training data behind
satgeze/Qwen3.5-0.8B-DSpark: public
prompts answered by the target model itself, so a drafter trains on precisely the distribution
it will draft for at inference. The unique part is the responses; they were generated by
Qwen3.5-0.8B and exist nowhere upstream.
Provenance, exactly
part
source
license
prompts
mlabonne/open-perfectblend
Apache-2.0
responses… See the full description on the dataset page: https://huggingface.co/datasets/satgeze/Qwen3.5-0.8B-selfdistill.ukr-tg-satire
Ukrainian Telegram satire & troll posts
A corpus of 92k posts from 17 public Telegram channels in the
satire/irony/parody register, harvested July 2026 via the public Telegram
API (history reaching back to 2018 for some channels). The channels are
Ukrainian-audience but the posts are a Ukrainian/Russian mix (the
main truha and sria_news feeds write mostly in Russian, the regional
truexa* branches mostly in Ukrainian), so the corpus carries
language: [ru, uk].
Two flavors are… See the full description on the dataset page: https://huggingface.co/datasets/hausmer/ukr-tg-satire.SATQuest
SATQuest Dataset
Paper: SATQuest: A Verifier for Logical Reasoning Evaluation and Reinforcement Fine-Tuning of LLMs
TL;DR. Synthetic CNF benchmark for LLM reasoning: 140 matched SAT/UNSAT pairs with n in [3, 16] and fixed ratio m=4n. The dataset stores only CNF formulas and solver stats; use the SATQuest Python library to render prompts/answers for SATDP, SATSP, MaxSAT, MCS, and MUS in four formats (math, DIMACS, story, dual story).
Data fields
id: unique identifier… See the full description on the dataset page: https://huggingface.co/datasets/sdpkjc/SATQuest.satsec-decomposition
SatSec Grounded Objective-Decomposition Dataset
Version 2.0 is a leakage-controlled replacement for the original dataset used in
A Controlled Candidate-Set Benchmark for Offline Satellite-Security Plan
Decomposition
(DOI 10.48550/arXiv.2607.26371).
It contains 24 authored full decompositions and 83 mechanically derived next-step rows
across 24 cases.
There are 82 train rows and 25 test rows; the six test cases never occur in training.
Important v2 correction
The… See the full description on the dataset page: https://huggingface.co/datasets/paolocmo/satsec-decomposition.BhojpuriCorpus
Bhojpuri Corpus (BhojpuriCorpus) — Monolingual Pretraining Dataset for Bhojpuri (bho)
Overview
BhojpuriCorpus is a monolingual pretraining dataset for the Bhojpuri language (ISO 639-3: bho), containing 386,032 documents and approximately 24.97 Million estimated tokens. It is compiled from multiple public sources and preprocessed for vocabulary training and language model pretraining.
Motivation
BhojpuriCorpus was compiled to aggregate, clean, and… See the full description on the dataset page: https://huggingface.co/datasets/Satyam810/BhojpuriCorpus.Chameleon-Radiology-Reportsukr-tg-satire-2
TG channels wave 2
A second scrape wave of 6 public Telegram channels
exported 2026-09-15 from their full histories via a
read-only MTProto API. Same schema as the main corpora
(hausmer/ukr-tg-media,
hausmer/ukr-tg-satire).
24,743 text posts across:
Foma_memes — Мемарня Sa-chan1917|ИзюмТГ (1,892 posts)
zelenskyi_vladimir — Владимир Зеленский (пародия) (612 posts)
karikaturnaya_satira — Карикатурная сатира (138 posts)
memarnya_rezerv — Ініціативна група «20 см» (1,169 posts)… See the full description on the dataset page: https://huggingface.co/datasets/hausmer/ukr-tg-satire-2.minif2f-satp-v4.27
minif2f-satp-v4.27
The 488-base-problem view of the miniF2F benchmark used by AlphaProof,
packaged for Lean 4.27.0 with initial proof goal_state values generated
for SATP v2.
This dataset is intended as the held-out evaluation (test) and
hyperparameter-tuning (validation) benchmark for SATP / sketch-and-prove
pipelines running in a Lean 4.27 environment.
Source
The single source of the problem statements is Google DeepMind
miniF2F@f0a20e1
(MiniF2F/Valid.lean +… See the full description on the dataset page: https://huggingface.co/datasets/ChristianZ97/minif2f-satp-v4.27.truha-news-satire
Truha news satire (Ukrainian)
A synthetic, LLM-generated dataset of Ukrainian "news" written in the
truha satirical register — short, meta-ironic, punchline-driven fake-news
items. The corpus was produced by distilling a target style (a Ukrainian
satirical news persona) into generated examples; no real user data is
included.
Intended use: style-transfer / imitation training for a Ukrainian satirical
news-writing assistant. Each example is a (system, instruction, output) triple.… See the full description on the dataset page: https://huggingface.co/datasets/hausmer/truha-news-satire.Qwen3.6-27B-selfdistill
Qwen3.6-27B self-distillation pairs (DSpark drafter training data)
The exact training data behind
satgeze/Qwen3.6-27B-DSpark (2.5-2.7x
measured decode speedup in llama.cpp): public prompts answered by the target model itself, so
the drafter trains on precisely the distribution it drafts for at inference. The responses were
generated by Qwen3.6-27B and exist nowhere upstream.
Provenance, exactly
part
source
license
prompts
mlabonne/open-perfectblend (12K… See the full description on the dataset page: https://huggingface.co/datasets/satgeze/Qwen3.6-27B-selfdistill.putnambench-satp
PutnamBench-SATP — Lean 4 (672 problems, SATP-normalized)
PutnamBench Lean 4 problems normalized for the SATP-DSP-Eval pipeline.
Upstream: trishullab/PutnamBench
(lean4/src/*.lean + informal/putnam.json, parsed via this repo's
datasets/putnam_bench/parse_lean4.py).
Differences from upstream
Field
Upstream
This repo
formal_statement
ends with := sorry
normalized to := by (no trailing tactic)
uuid
absent
sha256(canonical(formal_statement))[:16] after… See the full description on the dataset page: https://huggingface.co/datasets/ChristianZ97/putnambench-satp.SurgWound
Dataset Card for SurgWound
SurgWound is the first open-source dataset for surgical wound analysis across multiple procedure types.
SurgWound comprises 697 surgical wound images, each annotated by surgical experts at The Ohio State University Wexner Medical Center (OSWUMC).
Each image is accompanied by high-quality labels covering six surgical wound characteristic attributes and two diagnostic outcomes attributes.
SurgWound-Bench is the first multimodal benchmark for surgical… See the full description on the dataset page: https://huggingface.co/datasets/SATHEESHMA/SurgWound.proofnet-satp-v4.27
ProofNet#-SATP v4.27 — Lean 4 (371 problems)
ProofNet# (corrected reference
formalizations of ProofNet) normalized to the ChristianZ97/putnambench-satp-v4.27
schema for the SATP-DSP-Eval pipeline.
Provenance
Source: PAug/ProofNetSharp
@ a8da405fbd1e348a87445c2e562c747b7e26dc8f (MIT), 371 rows = 185 valid + 186 test.
Introduced in Reliable Evaluation and Benchmarks for Statement Autoformalization
(Poiroux, Weiss, Kunčak, Bosselut; EMNLP 2025 main;… See the full description on the dataset page: https://huggingface.co/datasets/ChristianZ97/proofnet-satp-v4.27.minif2f-satp
minif2f-satp
miniF2F benchmark with initial Lean 4 proof goal_state for both
test and validation splits. Intended as the held-out evaluation
(test) and hyperparameter-tuning (validation) sets for SATP /
sketch-and-prove pipelines trained on
NuminaMath-LEAN-satp.
Sibling datasets:
NuminaMath-LEAN-satp — main training set
NuminaMath-LEAN-satp-gaps — augmented train set with sub-goal records
NuminaMath-LEAN-satp-buffer— aesop-config replay buffer
Splits
from datasets… See the full description on the dataset page: https://huggingface.co/datasets/ChristianZ97/minif2f-satp.NuminaMath-LEAN-satp-gaps
NuminaMath-LEAN-satp-gaps
Lean 4 sub-goal (gap) dataset harvested from natural-language draft → Lean sketch → real-Lean goal-state extraction over the NuminaMath-LEAN formal statement pool. Each row is one open hole (sorry) inside a sketch, paired with the exact Lean goal-state at that hole, suitable as a per-sub-goal prove-step training signal.
This is the augmented training set complement to the entry-point
NuminaMath-LEAN-satp
main training set: where the main set carries one… See the full description on the dataset page: https://huggingface.co/datasets/ChristianZ97/NuminaMath-LEAN-satp-gaps.putnambench-satp-v4.27
putnambench-satp-v4.27
PutnamBench Lean 4 problems normalized for SATP evaluation, packaged for
Lean 4.27.0 with initial proof goal_state values.
Source
The single source is trishullab/PutnamBench@dc18909
(lean4/src/putnam_*.lean + informal/putnam.json); the upstream
lean4/lean-toolchain is leanprover/lean4:v4.27.0. All 672 Lean problems
are included and every row elaborates with zero errors under the pinned
toolchain.
Rows
672 (single train config… See the full description on the dataset page: https://huggingface.co/datasets/ChristianZ97/putnambench-satp-v4.27.NuminaMath-LEAN-satp-buffer-discard
NuminaMath-LEAN-satp-buffer
Aesop tactic configurations collected during SATP (Steering Aesop for
Theorem Proving) replay-buffer building, paired with the initial Lean
goal_state of each theorem. Each row is one
(theorem, aesop_config) → reward example, intended as positive /
negative replay material for training
SATP-aesop-policy.
Sibling datasets:
NuminaMath-LEAN-satp — main training set (formal_statement → goal_state)
NuminaMath-LEAN-satp-gaps — augmented train set with sub-goal… See the full description on the dataset page: https://huggingface.co/datasets/ChristianZ97/NuminaMath-LEAN-satp-buffer-discard.bharatschemes-v1
BharatSchemes Dataset v1
India's first comprehensive instruction-tuning dataset for government welfare schemes,
state programs, and corporate CSR/NGO initiatives.
Overview
This dataset contains 2029 instruction-following Q&A pairs covering:
Category
Coverage
Central Government Schemes
MyScheme.gov.in (3000+ schemes)
State Government Schemes
All 28 states + 8 UTs
Corporate CSR Programs
Tata, Infosys, Wipro, Reliance, Azim Premji, Adani, Mahindra
NGO… See the full description on the dataset page: https://huggingface.co/datasets/satyajitdas/bharatschemes-v1.verbal-confidence-saturation
Verbal Confidence Saturation Dataset
8,384 deterministic trials from a pre-registered study testing whether 3–9B instruction-tuned open-weight LLMs produce valid verbal confidence under minimal elicitation.
Paper: arXiv:2604.22215
Pre-registration: OSF
Code: GitHub
Dataset summary
Eight open-weight models were administered 524 TriviaQA items under numeric (0–100) and categorical (10-class) confidence elicitation with greedy decoding. All seven instruct models were… See the full description on the dataset page: https://huggingface.co/datasets/synthiumjp/verbal-confidence-saturation.proofnet-satp
ProofNet-SATP — Lean 4 (371 problems, SATP-normalized)
ProofNet (undergraduate-level mathematics theorem-proving benchmark, 371
problems from Rudin / Munkres / Dummit-Foote / Axler / Herstein /
Ireland-Rosen / Artin) normalized for the SATP-DSP-Eval pipeline.
Upstream: deepseek-ai/DeepSeek-Prover-V1.5
datasets/proofnet.jsonl — the Lean 4 port of
hoskinson-center/proofnet
shipped with DeepSeek-Prover V1.5.
Differences from upstream
Field
Upstream
This repo… See the full description on the dataset page: https://huggingface.co/datasets/ChristianZ97/proofnet-satp.NuminaMath-LEAN-satp-gaps-v4.27
NuminaMath-LEAN-satp-gaps-v4.27
Lean 4 sub-goal (gap) dataset harvested from natural-language draft → Lean
sketch → real-Lean goal-state extraction over the
NuminaMath-LEAN
formal statement pool, gated for Lean 4.27.0. Each row is one open
hole (sorry) inside a sketch, paired with the Lean goal-state recorded at
that hole during sketch extraction, suitable as a per-sub-goal prove-step
training signal.
This is the augmented-training complement to
NuminaMath-LEAN-satp-v4.27:
where… See the full description on the dataset page: https://huggingface.co/datasets/ChristianZ97/NuminaMath-LEAN-satp-gaps-v4.27.