datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
ToolACE
ToolACE
ToolACE is an automatic agentic pipeline designed to generate Accurate, Complex, and divErse tool-learning data.
ToolACE leverages a novel self-evolution synthesis process to curate a comprehensive API pool of 26,507 diverse APIs.
Dialogs are further generated through the interplay among multiple agents, guided by a formalized thinking process.
To ensure data accuracy, we implement a dual-layer verification system combining rule-based and model-based checks.
More details… See the full description on the dataset page: https://huggingface.co/datasets/Team-ACE/ToolACE.Scale-SWE
Immersion in the GitHub Universe: Scaling Coding Agents to Mastery
🔥 Highlights
Source from 6M+ pull requests and 23000+ repositories.
Cover 5200 Repositories.
100k high-quality instances.
71k trajectories from DeepSeek v3.2 with 3.5B token.
Strong performance: 64% in SWE-bench-Verified trained from Qwen3-30A3B-Instruct.
📣 News
2026-02-26 🚀 We released a portion of our data on Hugging Face. This release includes 20,000 SWE task… See the full description on the dataset page: https://huggingface.co/datasets/AweAI-Team/Scale-SWE.unclickbait-synthetic-27b-trajectories
Unclickbait Synthetic 27B Trajectories
Synthetic trajectory dataset generated by using rich structured JSON prompts and validated by two-stage judging pipeline.
Contents
: Full generated trajectories (current snapshot: 48,623 records out of 152,369 pristine event candidates).
: 30 benchmark test samples audited end-to-end through the 122B two-stage judge (Stage 1 integrity gate + Stage 2 4D scoring).
AM-DeepSeek-Distilled-40MFor more open-source datasets, models, and methodologies, please visit our GitHub repository and paper: DeepDistill: Enhancing LLM Reasoning Capabilities via Large-Scale Difficulty-Graded Data Training.
Due to certain constraints, we are only able to open-source a subset of the complete dataset.
Model Training Performance based on our complete dataset
On AIME 2024, our 72B model achieved a score of 79.2 using only supervised fine-tuning (SFT). The 32B model reached 75.8 and… See the full description on the dataset page: https://huggingface.co/datasets/a-m-team/AM-DeepSeek-Distilled-40M.CalibForge
CalibForge
CalibForge is a collection of 5,431 executable and verifiable terminal-agent tasks constructed with adversarial solver calibration.
CalibForge uses solver behavior during task construction in two complementary ways:
Multi-solver calibration retains tasks that expose disagreement across a heterogeneous solver pool.
Contrastive solver calibration targets a designated strong-pass and weak-fail capability relation.
Dataset structure
CalibForge/… See the full description on the dataset page: https://huggingface.co/datasets/AweAI-Team/CalibForge.BeyondSWE
BeyondSWE: Can Current Code Agent Survive Beyond Single-Repo Bug Fixing?
BeyondSWE is a comprehensive benchmark that evaluates code agents along two key dimensions — resolution scope and knowledge scope — moving beyond single-repo bug fixing into the real-world deep waters of software engineering.
✨ Highlights
500 real-world instances across 246 GitHub repositories, spanning four distinct task settings
Two-dimensional evaluation: simultaneously… See the full description on the dataset page: https://huggingface.co/datasets/AweAI-Team/BeyondSWE.skm-tea-mini
SKM-TEA Sample Data
This dataset consists of a subset of scans from the SKM-TEA dataset. It can be used to build tutorials / demos with the SKM-TEA dataset.
To access to the full dataset, please follow instructions on Github.
NOTE: This dataset subset should not be used for reporting/publishing metrics. All metrics should be computed on the full SKM-TEA test split.
Details
This mini dataset (~30GB) consists of 2 training scans, 1 validation scan, and 1 test scan from… See the full description on the dataset page: https://huggingface.co/datasets/arjundd/skm-tea-mini.moss-002-sft-data
Dataset Card for "moss-002-sft-data"
Dataset Summary
An open-source conversational dataset that was used to train MOSS-002. The user prompts are extended based on a small set of human-written seed prompts in a way similar to Self-Instruct. The AI responses are generated using text-davinci-003. The user prompts of en_harmlessness are from Anthropic red teaming data.
Data Splits
name
# samples
en_helpfulness.json
419049
en_honesty.json
112580… See the full description on the dataset page: https://huggingface.co/datasets/OpenMOSS-Team/moss-002-sft-data.FutureOmni
FutureOmni: Evaluating Future Forecasting from Omni-Modal Context for Multimodal LLMs
Predicting the future requires listening as well as seeing.
📖 Dataset Summary
Although Multimodal Large Language Models (MLLMs) demonstrate strong omni-modal perception, their ability to forecast future events from audio–visual cues remains largely unexplored, as existing benchmarks focus mainly on retrospective understanding.
FutureOmni is the first benchmark designed… See the full description on the dataset page: https://huggingface.co/datasets/OpenMOSS-Team/FutureOmni.AweAgent-Meta-NL2Repo
AweAgent-Meta-NL2Repo
This dataset provides the metadata used by AweAgent to run the NL2RepoBench evaluation.
If you are looking for the underlying benchmark itself (task design, repositories, test suites), please refer to the original project: multimodal-art-projection/NL2RepoBench.
Purpose
The AweAgent repository evaluates end-to-end repo-level code generation: given a natural-language project specification, the agent must produce a working Python package that… See the full description on the dataset page: https://huggingface.co/datasets/AweAI-Team/AweAgent-Meta-NL2Repo.gemma-2b-suite-explanations-residualgemma-2b-suite-maxacts-attn_out
Multi-doc-2025
Dataset Card for Multi-Doc-2025
Dataset Summary
Multi-Doc-2025 is a financial question-answering benchmark built from SEC Form 10-K annual reports of S&P 500 companies. It is designed to evaluate retrieval-augmented generation (RAG) and financial QA systems under three reasoning settings that are not jointly covered by existing financial QA benchmarks: cross-company reasoning, cross-year reasoning, and hybrid-modal reasoning over both text and tables.
The dataset… See the full description on the dataset page: https://huggingface.co/datasets/Anonymous-Team-HC-RAG/Multi-doc-2025.lexgraph
Lexgraph dataset
The data plane of Lexgraph —
German legislation modelled as Laws as Git: a temporal, multi-authority
event log with HEAD, commits, open/closed branches and evidence-bound merges
(Bund / Bayern / EU; Länder records only after verification at the originating
Landtag). Built 2026-07-19.
Git is a navigation metaphor, not a substitute for legal status. Every row's
official source and status controls whether it is current law, a pending branch
or a documented… See the full description on the dataset page: https://huggingface.co/datasets/SNTIQ-Team/lexgraph.gemma-2b-suite-maxacts-residual
teambench
TeamBench: Evaluating Agent Coordination under Enforced Role Separation
Overview
TeamBench is a benchmark of 851 task templates that expand to 931 seeded evaluation instances across 19 base categories (the leaderboard uses 21 refined categories; see paper §3.1). It evaluates whether LLM-based agent teams outperform a single oracle agent under OS-enforced role separation (Planner / Executor / Verifier in isolated sandboxes with distinct tool allow-lists), and… See the full description on the dataset page: https://huggingface.co/datasets/ybkim95/teambench.SciJudgeBench
SciJudgeBench Dataset
Training and evaluation data for scientific paper citation prediction, from the paper AI Can Learn Scientific Taste.
Given two academic papers (title, abstract, publication date), the task is to predict which paper has a higher citation count.
Resources: Project page, GitHub repository, SciJudge-4B-2605, and SciJudge-30B-2605.
Dataset Splits
Split
Examples
Description
train
720,341
Training preference pairs from arXiv papers
test… See the full description on the dataset page: https://huggingface.co/datasets/OpenMOSS-Team/SciJudgeBench.ABC-Bench
ABC-Bench
💻 Code |
📑 Paper |
📝 Blog
📖 Overview
ABC-Bench is a benchmark for Agentic Backend Coding. It evaluates whether code agents can explore real repositories, edit code, configure environments, deploy containerized services, and pass external end-to-end API tests (HTTP-based integration tests) across realistic backend stacks.
📊 Benchmark Composition
🚀 Why ABC-Bench?
End-to-End Lifecycle: repository exploration → code… See the full description on the dataset page: https://huggingface.co/datasets/OpenMOSS-Team/ABC-Bench.agentic_red_team
Agentic Red Team Tool-Calling Dataset
A multi-turn, tool-calling cybersecurity dataset where each example is a complete agentic trajectory — a realistic sequence of tool calls, tool responses, and reasoning steps that an AI agent would execute during an authorized red team engagement.
Overview
This dataset contains 5,000 agentic tool-calling examples across 20 offensive security sectors. Unlike traditional Q&A datasets, each row is a complete multi-turn trajectory… See the full description on the dataset page: https://huggingface.co/datasets/sh111111111111111/agentic_red_team.RED_team_tactics_dataset
Red Team Tactics
Overview
This dataset is a curated collection of advanced Red Team tactics designed for offensive cybersecurity operations at a DARPA-caliber standard.
It encompasses sophisticated techniques for cloud exploitation, browser-based attacks, zero-day vulnerabilities, and data exfiltration, aligned with MITRE ATT&CK techniques. The dataset is intended for training AI models, conducting Red Team simulations, or developing defensive countermeasures.… See the full description on the dataset page: https://huggingface.co/datasets/darkknight25/RED_team_tactics_dataset.cfever
CFEVER-data
Introduction to CFEVER
This repository contains the dataset for our AAAI 2024 paper, "CFEVER: A Chinese Fact Extraction and VERification Dataset". Paper link.
Leaderboard website
Please visit https://ikmlab.github.io/CFEVER to check the leaderboard of CFEVER.
How to load CFEVER
from datasets import load_dataset
# Get ready data ["train", "dev", "test"]
ds = load_dataset("IKMLab-team/cfever")
# Get separted data by:
train, dev, test =… See the full description on the dataset page: https://huggingface.co/datasets/IKMLab-team/cfever.Valen-Eval-General-5k
Valen-Eval-General-5k
GitHub · 中文 README · Preview model · Technical notes
✨ Introduction
Valen brings visual perception to System One decision-making: text, images or video in, candidate probabilities out. This dataset provides 5,000 image-based decision records for held-out evaluation, spanning visual question answering, interfaces, games and documents.
Each record contains one decision question, a target probability distribution, local image… See the full description on the dataset page: https://huggingface.co/datasets/Valen-Team/Valen-Eval-General-5k.DeepSeek-V4-Flash-0731-Teacher-Distillation-40513x
DeepSeek V4 Flash 0731 Teacher Distillation — 40,513 Retained Rows
Teacher-distillation corpus generated with
deepseek-ai/DeepSeek-V4-Flash-0731.
The original manifest contained 45,000 unique seeds.
Following generation, QC, retry-based repair, quarantine auditing,
and recovery adjudication, 40,513 rows were retained.
Composition
Bucket
Rows
Coding
5,601
Agentic
9,982
Cyber blue
13,000
Controlled cyber red
6,999
Tool use
4,931
Total
40,513… See the full description on the dataset page: https://huggingface.co/datasets/trjxter/DeepSeek-V4-Flash-0731-Teacher-Distillation-40513x.Valen-Eval-Game
Valen-Eval-Game
GitHub · 中文 README · Preview model · Technical notes
✨ Introduction
Valen brings visual perception to System One decision-making: text, images or video in, candidate probabilities out. This dataset supplies Sokoban screenshots and solver-derived action targets for training and evaluating visual decisions.
It contains 30,000 single-step training examples from 3,100 layouts, plus 100 complete-game evaluation levels and 500 single-step… See the full description on the dataset page: https://huggingface.co/datasets/Valen-Team/Valen-Eval-Game.gemma-2b-suite-maxacts-transcoder
teasecorpus
teasecorpus
A Chinese ChatML SFT dataset generated from the 「擅长捉弄的高木同学」Fandom wiki, with record-level contributor provenance tracked by originblame.
Summary: 1,410 question-answer pairs in ChatML format, covering 7 content types (chapters, characters, episodes, music, volumes, seasons, movies). Each record is traceable to its source wiki page and all contributors who edited that page, via originblame — a record-level provenance system. If a contributor requests content removal… See the full description on the dataset page: https://huggingface.co/datasets/tzbkk/teasecorpus.DIVE-SFT-20Knrftw-boss-arena
NRFTW Boss Arena — 同步视频 + 20Hz 遥测
📌 勘误（2026-08-31）
初次发布时本数据集包含 13 段视频，其中 6 段实际是纯黑画面已被移除。
原因：OBS 的游戏捕获在 2026-08-29 通宵批次没有挂上钩子，录下的是全黑帧。
当时用「文件体积在增长」判断录制正常 —— 这个判据是错的：NVENC 走固定码率，
纯黑同样会被填满到目标码率（30 分钟纯黑照样 2.1 GB）。
这 6 个 session 的日志完整可用，已在 metadata.jsonl 里标记为 quality: LOGS_ONLY。
现有 7 段视频均已逐个抽帧验证非黑。
No Rest for the Wicked 中 13 场 BOSS 战（其中 7 场带录像），每一帧画面都与同一时刻的模拟状态和玩家输入对齐。
全部战斗发生在同一个固定竞技场，由一个自动化 BOT 完成，因此 BOSS 行为之外的变量被刻意压到最小。
Telemetry from 13 boss fights (7 with… See the full description on the dataset page: https://huggingface.co/datasets/teawhite/nrftw-boss-arena.ru-reranker-teacher-scores
Teacher score cache for Russian reranker distillation
Raw logits of BAAI/bge-reranker-v2-m3 for ~5.4M (query, passage) pairs, grouped
by query (1 "pos" slot + candidates), produced for the ru-reranker-edge project.
teacher_scores.jsonl — one line per group: {"qid", "pos_score", "cand_scores": [...]} (raw logits, no normalization).
groups_pool_v5.jsonl — the matching texts: {"qid", "query", "pos_id", "pos_text", "cand_ids", "cand_texts"}.
Sources of queries/candidates:… See the full description on the dataset page: https://huggingface.co/datasets/sshalimov04/ru-reranker-teacher-scores.teacher-notes-severity
teacher-notes-severity
Synthetic dataset for training a local student-notes severity classifier (fine-tuned from Qwen/Qwen3-1.7B).
Each row is a chat-formatted prompt/completion pair: the user turn is a teacher note (single, compound,
or a cumulative running log), the assistant turn is a JSON label
{"category": "commendation|misbehavior|academic_concern", "severity": <int>, "escalate": <bool>}.
Severity scale: commendations are negative (-95..-10); routine notes 5-55; serious… See the full description on the dataset page: https://huggingface.co/datasets/jeremierostan/teacher-notes-severity.