datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
Terminal-Bench-Hard
Terminal-Bench Hard
Terminal-Bench Hard is a set of 100 challenging terminal-based agent tasks.
The tasks cover software engineering, debugging, data processing, system
administration, security, scientific computing, and related command-line
workflows.
Contents
tasks/: runnable tasks in Harbor format.
metadata/tasks.parquet: searchable task metadata and instructions.
Each task directory contains task.toml, instruction.md, an
environment/ directory, and verifier… See the full description on the dataset page: https://huggingface.co/datasets/Zhongzhi1228/Terminal-Bench-Hard.Recursive-Task-Synthesis
Recursive Task Synthesis
This dataset contains 37,484 validated command-line task instances produced
through recursive task synthesis. Public identifiers are opaque and stable.
metadata/tasks.parquet: one searchable row per task instance.
metadata/shard_manifest.jsonl: TAR sizes and SHA256 checksums.
data/tasks-*.tar: sanitized runnable task packages.
The searchable task rows include:
instruction: contents of instruction.md.
task_toml: contents of task.toml.
solution:… See the full description on the dataset page: https://huggingface.co/datasets/Zhongzhi1228/Recursive-Task-Synthesis.Recursive-Task-Synthesis-Trajectories
Recursive Task Synthesis Trajectories
This dataset contains 327,189 completed agent trajectories collected on
recursively synthesized command-line tasks. Public identifiers are opaque and
stable.
The trajectory JSON retains messages, actions, observations, and token counts.
Token-level log-probability arrays and duplicated debug/session captures are
excluded from the public packages.
metadata/trajectories.parquet: searchable trajectory metadata.
metadata/shard_manifest.jsonl:… See the full description on the dataset page: https://huggingface.co/datasets/Zhongzhi1228/Recursive-Task-Synthesis-Trajectories.minecraft-skins-1.1m-prerenderedhitabannotations_creators:
crowdsourced
language_creators:
crowdsourced
languages:
en
multilinguality:
monolingual
size_categories:
100K<n<1M
source_datasets:
original
task_categories:
tableqa, data2text
task_ids:
tableqa
zhoblimp
ZhoBLIMP Dataset
This is the Chinese version of the BLIMP (Benchmark of Linguistic Minimal Pairs) dataset.
Dataset Structure
The dataset contains 118 subtasks, each testing different linguistic phenomena in Chinese.
Usage
from datasets import load_dataset
# Load a specific subtask
dataset = load_dataset("your_username/zhoblimp", "BA_verb_le_b")
# Load all configs
all_configs = load_dataset("your_username/zhoblimp", "all")
Subtasks
The dataset… See the full description on the dataset page: https://huggingface.co/datasets/Junrui1202/zhoblimp.CoG-Evaluation-Data
CoG Evaluation Data
The seven evaluation sets used in CoG (EMNLP 2026), comprising 5,174 questions for knowledge-intensive QA across KG-based and text-based benchmarks. This collection supports evaluation of multi-hop retrieval and reasoning with graph, text, and hybrid RAG methods.
Paper · Code
Datasets
Dataset / HF config
Knowledge source
Questions
Raw file
KGQAGen / kgqagen
KG
1,079
KGQAGen-10k.json
CWQ / cwq
KG
1,024
cwq.json
QALD10-en /… See the full description on the dataset page: https://huggingface.co/datasets/zhougengxian/CoG-Evaluation-Data.BLINK_RotationQA_90degreeeval_policy_recordingThis dataset was created using LeRobot.
Dataset Structure
meta/info.json:
{
"codebase_version": "v3.0",
"robot_type": "bi_so100_follower",
"total_episodes": 118,
"total_frames": 127547,
"total_tasks": 6,
"chunks_size": 1000,
"data_files_size_in_mb": 100,
"video_files_size_in_mb": 500,
"fps": 30,
"splits": {
"train": "0:118"
},
"data_path": "data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet",
"video_path":… See the full description on the dataset page: https://huggingface.co/datasets/zhoumiaosen/eval_policy_recording.bag_grocery_humanRecursive-Task-Synthesis-Quality-1K
Recursive Task Synthesis Quality 1K
This dataset contains 1,000 quality-selected, validated command-line task
instances. It is a curated subset of the
Recursive Task Synthesis dataset.
Public task and group identifiers are opaque and stable across both datasets.
Selection
The subset was selected from 37,484 validated tasks using structural and safety
checks, two-pass semantic review, strict gates for instruction clarity,
instruction-verifier alignment, verifier… See the full description on the dataset page: https://huggingface.co/datasets/Zhongzhi1228/Recursive-Task-Synthesis-Quality-1K.openai_MMMLU_zhotwo_clublichess-2500plus-gamesThis dataset contains all games that was played by two 2500+ elo players, and ended with a checkmate.
chess-reasoning-cot-evalsSAMTok_Training_DataRotationQAColorBench
🎨 ColorBench
📖 Paper | 💻 GitHub
ColorBench is a multimodal dataset to comprehensively assess capabilities of VLMs in color understanding, including color perception, reasoning, and robustness, introduced in "ColorBench: Can VLMs See and Understand the Colorful World? A Comprehensive Benchmark for Color Perception, Reasoning, and Robustness".
It provides:
More than 5,800 image-text questions covering diverse application scenarios and practical challenges for VLMs evaluation.
3… See the full description on the dataset page: https://huggingface.co/datasets/umd-zhou-lab/ColorBench.makise-kurisu-vn-voicelines
Makise Kurisu VN dialogue
Transcribed using Whisper Large-V2 from this video.
Clips were separated via pydub, so some text may be incorrect. I have not cleaned it up at all.
Intended for TTS model training.
I do not own any of the content.
2Arms_One_Object_ShipThis dataset was created using LeRobot.
Dataset Structure
meta/info.json:
{
"codebase_version": "v3.0",
"robot_type": "bi_so100_follower",
"total_episodes": 20,
"total_frames": 25436,
"total_tasks": 1,
"chunks_size": 1000,
"data_files_size_in_mb": 100,
"video_files_size_in_mb": 500,
"fps": 30,
"splits": {
"train": "0:20"
},
"data_path": "data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet",
"video_path":… See the full description on the dataset page: https://huggingface.co/datasets/zhoumiaosen/2Arms_One_Object_Ship.landuse_50000BLINK_PositionQA_augmentedcloth_foldingThis dataset was created using LeRobot.
Dataset Structure
meta/info.json:
{
"codebase_version": "v3.0",
"robot_type": "bi_so100_follower",
"total_episodes": 8,
"total_frames": 13684,
"total_tasks": 1,
"chunks_size": 1000,
"data_files_size_in_mb": 100,
"video_files_size_in_mb": 500,
"fps": 30,
"splits": {
"train": "0:8"
},
"data_path": "data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet",
"video_path":… See the full description on the dataset page: https://huggingface.co/datasets/zhoumiaosen/cloth_folding.portable-hagridv2-mediapipe-hand
portable_hagridv2_mediapipe_hand
这是一个从 HaGRIDv2 派生出来的便携版 MediaPipe Hand 测试数据集，用于 palm detection 和 hand landmark 验证。
它是更大的 hagridv2_512px_mediapipe_hand 生成数据集的过滤子集，体量更小，方便随 ONNX / Ascend 310B 运行时验证工程一起迁移。
数据集包含两类 MediaPipe 风格的目标：
palm_detection：单一 palm 类，包含 palm_bbox_xyxy 和 palm7_keypoints
hand_landmark：同一只手的 full21_keypoints
数据集也保留了原始 HaGRIDv2 手势标签，作为辅助来源标签。这个手势标签不是 palm detector 的类别。
数据集概览
图片总数：9754
数据划分：train、valid、test
划分数量：train=7246，valid=845，test=1663… See the full description on the dataset page: https://huggingface.co/datasets/zhouxzh/portable-hagridv2-mediapipe-hand.ChartAlignBench
📊 ChartAlignBench
📖 Paper | 💻 GitHub
ChartAlignBench is a multi-modal benchmark designed to evaluate vision-language models (VLMs) on dense-level chart grounding and multi-chart alignment to comprehensively assess fine-grained chart understanding in VLMs.
🌐 Overview
ChartAlignBench contains 9K+ instances, divided into three evaluation subsets:-
Data Grounding & Alignment: Paired charts differ in underlying data values visualized by the chart.
Attribute Grounding &… See the full description on the dataset page: https://huggingface.co/datasets/umd-zhou-lab/ChartAlignBench.hopton2026speech
hopton2026speech — derived data
Audio-free derived artifacts behind Estimating Surprisal from the Speech Signal (Hopton et al., 2026), restricted to RSpin lists 3–6 (the analysed set). No audio, transcript-of-record, or waveform-reconstructable content is included.
Built by pipeline/hf_dataset/build_hf_dataset.py in the study repo. The aggregate probability tables the study reads directly are committed in the repo under pipeline/inputs/; this dataset holds the per-sample… See the full description on the dataset page: https://huggingface.co/datasets/zhopto3/hopton2026speech.toolgrad-500
🛠️ ToolGrad-500: A Tool-Use Dataset Generated by ToolGrad (ACL 2026 Findings)
ToolGrad-500 is a synthetic tool-use dataset targeting single-turn parallel function calling scenarios. The dataset is generated using the ToolGrad framework (ACL 2026 Findings), leveraging textual gradients to optimize tool selection and calling.
📊 Dataset Structure
This dataset consists of formatted single-turn conversations suitable for standard conversational trainers.… See the full description on the dataset page: https://huggingface.co/datasets/zhongyi-zhou/toolgrad-500.Bespoke-Stratos-17k-Train-Posterior-PAretinaface_widerface
retinaface_widerface
这个目录用于生成可直接上传到 Hugging Face Datasets 的 WiderFace 教学版数据。
目标格式
转换完成后会生成两个 split：
train
val
对应文件路径默认是：
train/train-00000-of-00001.parquet
val/val-00000-of-00001.parquet
数据来源
train 来自 data/widerface/train/label.txt 和 data/widerface/train/images
val 来自 data/widerface/val/images 和 widerface_evaluate/ground_truth 下的 4 个 mat 文件
运行方式
在仓库根目录执行：
conda run -n retinaface python data/retinaface_widerface/build_parquet.py
如果环境里还没有… See the full description on the dataset page: https://huggingface.co/datasets/zhouxzh/retinaface_widerface.Variational-Posterior-PB-7B-GML-mix