datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
Multi-SWE-RL-Verified
Multi-SWE-RL-Verified
Gold-patch-validated subset of
PrimeIntellect/Multi-SWE-RL-Reupload
(ByteDance's Multi-SWE-RL): 2,232 / 4,703 rows across
C, Go, Java, JavaScript, Rust, and TypeScript that produce a clean reward signal end-to-end.
Default dataset of the multiswe_v1 taskset.
Changes vs upstream
Starting from the 4,703-row re-upload:
C++ dropped wholesale — 0/449 rows passed gold-patch validation in pass 1; the images are
broken for scoring, not merely… See the full description on the dataset page: https://huggingface.co/datasets/PrimeIntellect/Multi-SWE-RL-Verified.epoch_ai_swebench_verified
Epoch AI SWE-bench Verified Traces
Complete public trace archives and an analysis-ready Parquet conversion of Epoch AI's SWE-bench Verified evaluations.
Contents
34 published evaluation runs covering 16,456 traces (484 SWE-bench instances per run).
data/: loadable Parquet data, one exact trace per row.
original/: the byte-identical .eval archives published by Epoch AI.
run_metadata/: non-sample files from each .eval archive (header.json, summaries, reductions… See the full description on the dataset page: https://huggingface.co/datasets/naderalfares/epoch_ai_swebench_verified.R2E-Gym-Subset-Verified
R2E-Gym-Subset-Verified
Gold-patch-validated subset of
R2E-Gym/R2E-Gym-Subset
(paper). The train split contains
4,522 / 4,578 rows (98.78%) verified scoreable end-to-end: apply the gold patch, run the
upstream /testbed/run_tests.sh baked into the row's image, check the parsed outcomes against
expected_output_json.
Changes vs upstream
Validation-only subset — our passes, run in fresh sandboxes per row: one full pass at
concurrency 200, then a 10× retry pass over… See the full description on the dataset page: https://huggingface.co/datasets/PrimeIntellect/R2E-Gym-Subset-Verified.swebench-verified-deepseek-v4-flash-failure-analysis
SWE-bench Verified runs & failure analysis — DeepSeek-V4-flash (local) × mini-swe-agent
Per-instance analysis of SWE-bench Verified runs of a locally-served DeepSeek-V4-flash model
driven by mini-swe-agent, graded with the official
SWE-bench harness. Each instance carries the full agent trajectory, a readable transcript, the
submitted patch, the harness test output, deterministic metrics, and a hand-verified qualitative
root-cause diagnosis.
Current numbers (resolve rates… See the full description on the dataset page: https://huggingface.co/datasets/daaain/swebench-verified-deepseek-v4-flash-failure-analysis.AIME24-25_CoT_Verification
Dataset for ICLR 2026 Paper: Reinforcement Learning with Verifiable Rewards Implicitly Incentivizes Correct Reasoning in Base LLMs
📌 Dataset Summary
This dataset contains the rollouts (reasoning traces) and verification results used in our ICLR 2026 paper. The data allows for the analysis of how Reinforcement Learning with Verifiable Rewards (RLVR) incentivizes the correct reasoning of Large Language Models (LLMs) on challenging mathematics benchmarks.
The dataset… See the full description on the dataset page: https://huggingface.co/datasets/XumengWen/AIME24-25_CoT_Verification.agent-trajectories-swe-bench-test-minus-verified
Agent Trajectories: SWE-bench Test \ Verified — Mixed Teachers (gpt-5.2 / gpt-5-mini)
Summary
Full multi-turn agent trajectories collected from the SWE-bench Test minus Verified split
(i.e., SWE-bench Test instances that are not part of SWE-bench Verified).
Intended for SFT of agent models on coding tasks.
Data Collection
Each trajectory was produced by a GT-aware lookahead agent that, at every turn:
Sampled a candidate response from both gpt-5.2 and… See the full description on the dataset page: https://huggingface.co/datasets/JetBrains-Research/agent-trajectories-swe-bench-test-minus-verified.verified-research-reasoning-trajectories
Verified Research Reasoning Trajectories for RLVR
This repository is the public sample and schema repository for Ulam's research-level mathematical reasoning trajectories for reinforcement learning with verifiable rewards (RLVR), process supervision, judge training, proof criticism, and private evaluations.
Ulam Verified Research Reasoning Trajectories are proof-process data for RLVR. Each record contains a normalized research problem, a golden or partial-golden proof graph… See the full description on the dataset page: https://huggingface.co/datasets/ulamai/verified-research-reasoning-trajectories.veri-bilimci-insight-diyalog-tr-16.2k
🇹🇷 Veri Bilimci Insight Diyalog Veri Seti (TR, 16.2K) — %100 Türkçe Metin
Gerçek dünya blogları, uzman soru-cevap içerikleri ve akademik makale metinlerinden üretilmiş; veri madenciliği ve uygulamalı veri bilimi karar diline odaklanan, %100 Türkçe çok turlu diyalog veri seti.
🧠 Bu Veri Seti Ne Amaçla Üretildi?
Amaç, modeli teorik tanım ezberinden çıkarıp bağlama göre karar veren veri bilimci davranışına yaklaştırmaktır.
Her örnekte yöntem seçimi, alternatif kıyası… See the full description on the dataset page: https://huggingface.co/datasets/zero9tech/veri-bilimci-insight-diyalog-tr-16.2k.SWEbench-Verified-eval150-M2.7-Qwen3.5-9B-orch-7arms-2repeats-w32-20260920
SWE-bench Verified eval150 — M2.7 × Qwen3.5-9B, seven arms, two repeats, 32 concurrency
Campaign 2026-09-20. 14/14 independent full150 runs audited. Complete accuracy evidence.
Evaluation mode is orch: MiniMax-M2.7 orchestrator and the specified Qwen3.5-9B worker. Training mode is labeled independently. All runs use 32 concurrent episodes, 10GiB Docker sandboxes, four TP1 workers and one TP4/EP4 coordinator. Frozen regression-gated prompts, decoding and canonical verifier match… See the full description on the dataset page: https://huggingface.co/datasets/CharlieLLL/SWEbench-Verified-eval150-M2.7-Qwen3.5-9B-orch-7arms-2repeats-w32-20260920.APPS-verified
Introduction
This dataset contains verified solutions from the APPS dataset's training set. Solutions that fail to pass all the test cases are removed. Problems with no correct solution are also removed.
The solutions were executed on Intel E5-2620 v3 CPUs with the execution timeout set to 10 seconds.
Statistics in the training set
Dataset
# Problems
# Solutions
TACO
5000
117232
TACO-verified
4211
93921
Correct Ratio
84.22%
80.12%
sci-agent-verification-cascade
Scientific Agent Verification Cascade
Public evaluation fixtures and verified aggregate results for testing whether
scientific claims keep their source, meaning, uncertainty, and verification
requirements as they move between AI agents.
This dataset accompanies the
Scientific Agent Verification Cascade
codebase. Version 0.2.0
contains synthetic evaluation data and aggregate-only results. It contains no
raw hosted-model response, private holdout identifier,
source-record… See the full description on the dataset page: https://huggingface.co/datasets/jang1563/sci-agent-verification-cascade.tmax-tasks-selfgen-qwen35-9b-20260919-1k-verified
tmax self-generated tasks — Qwen3.5-9B (verified arm)
The 1,042 tasks from
…-20260919-1k,
each graded against the issue-#12 rubric by the same model that generated them
(hamishivi/Qwen3.5-9B). Using the generator as its own reviewer is deliberate: the
question is whether an open-weights model can carry both halves of the loop. A stronger
reviewer would answer a different question.
The grader sees instruction / setup.sh / tests only. truth is withheld from it,
so it is no better… See the full description on the dataset page: https://huggingface.co/datasets/osieosie/tmax-tasks-selfgen-qwen35-9b-20260919-1k-verified.swerebench-traces-raw-source-verification-enhanced-20260617
SWE-rebench Raw Source Verification Enhanced 20260617
This is a private raw source dataset for building refined mini-swe-agent SFT datasets. It is intentionally not tokenized and intentionally preserves source data plus metadata for downstream filtering, masking, weighting, and audit. Do not treat every row as a clean endpoint solve.
Download
The full dataset directory is uploaded as a single compressed archive:
hf download… See the full description on the dataset page: https://huggingface.co/datasets/eewer/swerebench-traces-raw-source-verification-enhanced-20260617.devstral-swe-bench-verified
Devstral SWE-bench Verified Trajectories
This dataset contains 4,000 mini-SWE-agent trajectories generated by
mistralai/Devstral-Small-2-24B-Instruct-2512 on a fixed 250-problem training split of
SWE-bench Verified. There are 16 independently seeded rollouts per problem.
Every row preserves the exact request prefix mini-SWE-agent supplied before Devstral's
first generation, the complete ordered rollout, a sanitized final patch, and termination
metadata. Official SWE-bench… See the full description on the dataset page: https://huggingface.co/datasets/suryadv/devstral-swe-bench-verified.ReForm-DafnyComp-Benchmark
Re:Form Datasets
This repository contains the datasets used in the paper Re:Form -- Reducing Human Priors in Scalable Formal Software Verification with RL in LLMs: A Preliminary Study on Dafny.
The Re:Form project introduces a framework for code to specification generation using large language models, based on Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL). This work systematically explores ways to reduce human priors in scalable formal software verification by… See the full description on the dataset page: https://huggingface.co/datasets/Veri-Code/ReForm-DafnyComp-Benchmark.vericoding
Vericoding
A benchmark for vericoding: formally verified program synthesis
Sergiu Bursuc, Theodore Ehrenborg, Shaowei Lin, Lacramioara Astefanoaei, Ionel Emilian Chiosa, Jure Kukovec, Alok Singh, Oliver Butterley, Adem Bizid, Quinn Dougherty, Miranda Zhao, Max Tan, Max Tegmark
We present and test the largest benchmark for vericoding, LLM-generation of formally verified code from formal specifications - in contrast to vibe coding, which generates potentially buggy code from a… See the full description on the dataset page: https://huggingface.co/datasets/beneficial-ai-foundation/vericoding.BIRD-Verified-CoT-2462-GPT5.4
BIRD-Verified-CoT-2462 (GPT-5.4 distilled)
Likely the first publicly available CoT-augmented Text-to-SQL dataset built on top of expert-verified BIRD data.
This dataset combines two state-of-the-art ingredients:
ReViSQL's BIRD-Verified subset — 2,462 SQL-expert verified examples (multi-round review by UIUC team), eliminating the ~50% annotation noise of the original BIRD train set.
GPT-5.4 (via Codex CLI) — distilled into structured 6-section Chain-of-Thought traces using… See the full description on the dataset page: https://huggingface.co/datasets/wenyupapa/BIRD-Verified-CoT-2462-GPT5.4.swebench-verified-kimi-k2p6-traces
SWE-bench Verified Kimi K2.6 Reasoning Traces
This dataset contains reasoning traces generated on princeton-nlp/SWE-bench_Verified using fireworks_ai/kimi-k2p6-high with a mini-swe-agent based harness. It is intended for research and distillation of software-engineering agents.
The repository is published with three configs because each table has a different schema:
raw_trajectories: one row per SWE-bench instance with the patch, sanitized result JSON, full trajectory JSON, message… See the full description on the dataset page: https://huggingface.co/datasets/MemoryAsModality/swebench-verified-kimi-k2p6-traces.regexgym-verified-traces
RegexGym-Verified-Traces
Reasoning traces for writing regexes from examples. Each record shows a task (some strings that
should match, some that shouldn't), the teacher's chain-of-thought, and the regex it landed on.
Every trace here actually solved the task's hidden holdout — the regex was run against
examples the teacher never saw, and only exact solves were kept. The ground-truth regexes aren't
in the released records; the model has to earn its answer.
What's in… See the full description on the dataset page: https://huggingface.co/datasets/ctokx/regexgym-verified-traces.numeric-claim-verifier
Numeric Claim Verifier (Science) — Adaption AutoScientist
Programmatically verified prompt/completion pairs for scientific and statistical numeric claim verification.
Labels
correct — claim matches ground-truth tables
wrong_direction — trend/sign reversed
wrong_magnitude — right direction, wrong size (25–70% offset)
unverifiable — no matching source row (real entity + absent metric)
Sources
Our World in Data CO₂ / Energy
WHO GHO life expectancy… See the full description on the dataset page: https://huggingface.co/datasets/mishface123/numeric-claim-verifier.Z3-Verified-Reasoning-Graphs
Z3-Verified Constraint Reasoning Dataset
5k Baseline · Production-Ready · Zero Label Noise
The Problem This Solves
Most synthetic reasoning datasets only show the "happy path". Real reasoning requires knowing when to backtrack.
Open-source LLMs hallucinate on constraint satisfaction problems because they are trained on fluent-sounding but logically inconsistent traces. This dataset is different:
❌ No LLM-generated reasoning — zero hallucinations, zero label noise
✅… See the full description on the dataset page: https://huggingface.co/datasets/nagygabor/Z3-Verified-Reasoning-Graphs.brainblast-verified-footgun-corpus
Brainblast — Verified SDK Footgun Corpus (free sample)
The only code-training data that ships with a machine-checkable proof. Each
record is a real insecure→fixed code footgun with a replayable RED→GREEN
receipt: a deterministic checker fails the insecure version and passes the fixed
one. You don't trust the labels — you replay the proof.
This repo is a free 40-record sample (receipt-only tier). The full corpus is
4,183 proven records across 154 SDKs and 9 vulnerability classes… See the full description on the dataset page: https://huggingface.co/datasets/dsb117/brainblast-verified-footgun-corpus.retailopt-10k-verified
RetailOpt-10K Verified
Solver-verified expansion of Jacoblian/RetailOpt-190.
Overview
Property
Value
Base scenarios
190
Variations per base
53
Target total
10,070
Solver
Pyomo + HiGHS
Verification
Optimal objective per scenario
Variation Axes
Each base scenario is perturbed along:
demand_scale — scale demand curves (0.7–1.4×)
capacity_scale — scale production and cold storage capacity (0.6–1.3×)
cost_multiplier — scale… See the full description on the dataset page: https://huggingface.co/datasets/alirezaaminzadeh/retailopt-10k-verified.verified-analytics-tasks
Verified Analytics Tasks
150+ mainly small analytics and data-engineering tasks. The point of the set is
the answer key: every task ships its own automated checker, and every gold
answer was run through that checker and scored a clean 1.0 before the task was
allowed in. So the labels are more like "here's the checker, score it
yourself" instead of "just trust me bro."
I wanted to create a synthetic dataset inspired by this paper:
Autodata: An agentic data scientist to create… See the full description on the dataset page: https://huggingface.co/datasets/Eve39570/verified-analytics-tasks.veri-bilimci-diyalog-8k-tr
🇹🇷 Veri Bilimci Diyalog Veri Seti 7.6K (TR) — %100 Türkçe Metin
Gerçek dünya blogları, uzman soru-cevap içerikleri ve akademik makale metinlerinden üretilmiş; veri madenciliği ve uygulamalı veri bilimi karar diline odaklanan, %100 Türkçe çok turlu diyalog veri seti.
📋 Veri Seti Özeti
Özellik
Değer
Dil
Türkçe (%100 metin saflığı)
Toplam kayıt
7,656
Split
train: 6,124 · test: 1,532
Kaynak dağılımı
makale: 5,475 · yazı: 2,181
Konuşma başına tur sayısı
6… See the full description on the dataset page: https://huggingface.co/datasets/murataksit34/veri-bilimci-diyalog-8k-tr.m9-verifier-38k-aligned
M9 Verifier 38K Aligned
This dataset contains 38,564 prompts with verifier-compatible gold answers for
an M9 RLVR-GRPO experiment in a unified post-training study with Qwen3-1.7B.
It is an independent research artifact, not an official release from the model
or paper authors.
The bank was reconstructed from the frozen
YangyiH/openreasoning_mixed_100k
prompt mixture. Every recovered row was matched to the frozen base row by
domain, source shard, and prompt SHA-256 before verifier… See the full description on the dataset page: https://huggingface.co/datasets/YangyiH/m9-verifier-38k-aligned.bullet-swebench-verified
Bullet on SWE-bench Verified — 479/500 = 95.8%
Results for the Bullet coding agent on all 500 instances of
SWE-bench Verified, graded by the official swebench.harness.run_evaluation
scorer. Every instance was attempted and graded; there are no empty patches.
479 / 500 resolved = 95.8%
Run with the Bullet harness on gpt-5.6-sol at high reasoning effort, one attempt per instance.
By repository
repository
resolved
django
223/231
96.5%
sympy
73/75
97.3%… See the full description on the dataset page: https://huggingface.co/datasets/daviddata1/bullet-swebench-verified.ceo-quotes-verified-sample
🎙️ CEO Transcripts — Verified Executive Interviews
The World's Largest Database of Verified C-Suite Transcripts
20,000+ Executives · 100,000+ Transcripts · 400,000+ Quotes · S&P 500 + NASDAQ + Global Leaders
🔥 What's In This Sample?
This is a free evaluation sample from CEOInterviews.ai featuring 9 of the most market-moving voices in finance, tech, and policy.
Executive
Role
Why They Matter
Jensen Huang
CEO, NVIDIA
Every AI… See the full description on the dataset page: https://huggingface.co/datasets/codelucas/ceo-quotes-verified-sample.verifimind-peas-eval
VerifiMind-PEAS Evaluation Dataset
DOI: 10.5281/zenodo.21276884 · License: MIT · Version: 1.0
A human-annotated evaluation dataset for measuring the performance of the VerifiMind-PEAS multi-agent epistemic verification system. Ground-truth labels were assigned by a single domain-expert annotator whose final verdicts and confidence were human judgments; disclosed LLM comprehension assistance was used during annotation (see Annotation Protocol — transparency is a design commitment… See the full description on the dataset page: https://huggingface.co/datasets/YSenseAI/verifimind-peas-eval.spai-ss6-corpus-medical-o1-verifiable
SPAI SS6 Medical O1 Verifiable Thai Index
Index repo for the imported Thai medical verifiable-problem dataset config.
This is a lightweight index dataset repo. It does not duplicate the full corpus.
The full Parquet data lives in the canonical repository config below.
Canonical Data
Canonical repo: SPAISS6F1/spai-ss6-llm-1b-thai-corpus
Canonical config: medical_o1_verifiable_problem_thai
Rows in canonical config: 40,906
Parquet size in canonical config: 0.00 GB… See the full description on the dataset page: https://huggingface.co/datasets/SPAISS6F1/spai-ss6-corpus-medical-o1-verifiable.