datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
stack-v3-train
🥞 The Stack v3
What is it?
What is being released
How to download and use it
Dataset statistics
Dataset structure
Dataset creation
Considerations for using the data
Additional information
What is it?
The Stack v3 is the largest, most up-to-date open dataset of source code, crawled directly from GitHub and built to pre-train code LLMs with full-repository context. It is the successor to The Stack v2 and, like its predecessor, is released to make the training… See the full description on the dataset page: https://huggingface.co/datasets/HuggingFaceCode/stack-v3-train.agibot_alpha_v30This dataset was created using LeRobot.
Dataset Structure
meta/info.json:
{
"codebase_version": "v3.0",
"robot_type": "AgiBot_A2D",
"total_episodes": 28122,
"total_frames": 47613574,
"total_tasks": 30,
"chunks_size": 1000,
"data_files_size_in_mb": 100,
"video_files_size_in_mb": 500,
"fps": 30,
"splits": {
"train": "0:28122"
},
"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/cadene/agibot_alpha_v30.TxT360-v2
TxT360-v2
Dataset Description
Pre-training sources for the K2 Horizon training data release. This repository is part of the K2 Horizon collection.
The repository is organized into multiple subsets. Every subset has a train split backed by Parquet shards.
K2 Horizon Dataset Series
Dataset repository
Focus
Subsets
IFM/TxT360-v2
Web and question-answering text
3
IFM/Code-Reasoning
Code reasoning and task synthesis
7
IFM/Math-Reasoning… See the full description on the dataset page: https://huggingface.co/datasets/IFM/TxT360-v2.robotwin_3d
RoboTwin 2.0 — 3D (RGB + Depth)
Bimanual manipulation data from the RoboTwin 2.0 simulator, in LeRobot v2.1 format, with per-camera ground-truth depth alongside RGB.
Tasks
50
Episodes
27,500 (550 per task, contiguous)
Frames
6,183,813
Robot
ALOHA-style bimanual, 14-DoF
Control rate
50 Hz
Cameras
3 (cam_high, cam_left_wrist, cam_right_wrist)
Resolution
240 × 320
Language instructions
1,039,891 unique corpus-wide; 100 entries per episode
Total size… See the full description on the dataset page: https://huggingface.co/datasets/flex-pi/robotwin_3d.behaviour1k-Qwen3-features
BEHAVIOR-1K Qwen3 skill features
Per-frame conditioned features e_t = Phi(f_t, L_sub^(j), L), mean-pooled primitive skill latents S_j, aligned proprioception q_t, actions a_t, and subtask progress p_t.
These are the inputs and targets for a Primitive Skill Composer VLA Skill Predictor.
Ground-truth primitives come from BEHAVIOR-1K's hand-authored
primitive_annotation, so the segmentation is human-labelled rather than
predicted, and nothing here depends on a keyframe detector.… See the full description on the dataset page: https://huggingface.co/datasets/erl-hub/behaviour1k-Qwen3-features.Wan2.2-Syn-121x704x1280_32k
FastVideo Synthetic Wan2.2 720P dataset
FastVideo Team
Paper |
Github |
Project Page
Abstract
Scaling video diffusion transformers (DiTs) is limited by their quadratic 3D attention, even though most of the attention mass concentrates on a small subset of positions. We turn this observation into VSA, a trainable, hardware-efficient sparse attention that replaces full attention at \emph{both} training and inference. In VSA, a… See the full description on the dataset page: https://huggingface.co/datasets/Hahshshsshbs/Wan2.2-Syn-121x704x1280_32k.robocasa365-pretrain-mg
Pretraining (MimicGen) — atomic
MimicGen-generated rollouts across 60 atomic tasks (~10,000 demos/task). 1,615 hours total, generated by scripted augmentation from human demonstrations.
Part of the RoboCasa365 collection. Flat LeRobot v3.0 mirror of RoboCasa365 — standard layout, drop-in loadable.
Stats
Episodes: 536,030
Frames: 116,246,439 (20 fps → 1615 h)
Tasks: 720 (natural-language phrasings; underlying RoboCasa task classes: 60)
Cameras: 3 × 256×256 h264… See the full description on the dataset page: https://huggingface.co/datasets/ember-lab-berkeley/robocasa365-pretrain-mg.jailbreak-deepseek-v3.2-expgs-images-v3TrainingData_Stage3
AnchorSR Stage3 · metric-v1.0
直接选择 Small / Large
配置
训练题数
用途
small
1,000,000
先验证答案监督／先验恢复，按新版 Large 联合分布抽样
large
89,801,853
筛选后的完整训练集合，包含 Small 全部样本
from datasets import load_dataset
data = load_dataset('AnchorSR/TrainingData_Stage3', 'small', # 或 large
revision='metric-v1.0', streaming=True)
这是对 scaling-v1.0 的语义筛选与统一任务分类，不是增加新数据源。
Large 从 89,828,269 题保留 89,801,853 题，隔离 26,416 题。
旧标签 scaling-v1.0 / video-v1.0 / large-v1.0… See the full description on the dataset page: https://huggingface.co/datasets/AnchorSR/TrainingData_Stage3.VidaForge-3M
3.14 million scene-level video clips with multi-level captions, camera labels, semantic tags, quality signals, and duplicate groups.
Paper
·
VidaForge Code
·
Project Blog
·
Source Dataset
Overview
VidaForge-3M is a large-scale video pretraining dataset produced with
VidaForge, an open data pipeline for
building and studying video foundation model pretraining data. The pipeline and
dataset are described in the paper
VidaForge: Open Research Infrastructure… See the full description on the dataset page: https://huggingface.co/datasets/VidaForge/VidaForge-3M.blip3-kale
🥬 BLIP3-KALE:Knowledge Augmented Large-scale Dense Captions
BLIP3-KALE is an open-source dataset of 218 million image-text pairs, featuring knowledge-augmented dense captions combining web-scale knowledge with detailed image descriptions.
Paper: [To be added]
Uses
BLIP3-KALE is designed to facilitate research in multimodal pretraining. The dataset can be used for training large multimodal models that require factually grounded, dense image captions. It has already been an… See the full description on the dataset page: https://huggingface.co/datasets/Salesforce/blip3-kale.data_3
SpatialEncoder WDS release (in progress)
This repository contains a partition of spatialencoder-wds-native-v1, released
as uncompressed WebDataset tar shards, normally about 1 GiB. All five
repositories are parts of the same release; consult each manifest.json.
The manifest lists only uploaded shards whose remote size and SHA-256 have
been verified. An incomplete manifest is not a complete dataset.
New uploads use bucketed paths such as… See the full description on the dataset page: https://huggingface.co/datasets/xxxspatialencoderwds3/data_3.paired-llama-3.2-1b-embeddings-lmsys-chat-1m
Paired Llama 3.2 1B Token Embeddings (LMSYS-Chat-1M)
This dataset contains paired activations corresponding to single token locations extracted from Meta's Llama 3.2 1B Instruct on conversations from LMSYS-Chat-1M.
Embeddings are provided for layers 5 through 14, which capture the most interesting intermediate representations.
This dataset was built to study things like:
Learning different basis for activations at a given layer
Studying if there are cases where position encodes… See the full description on the dataset page: https://huggingface.co/datasets/scaleinvariant/paired-llama-3.2-1b-embeddings-lmsys-chat-1m.MRSDrama
ISDrama: Immersive Spatial Drama Generation through Multimodal Prompting
Yu Zhang*, Wenxiang Guo*, Changhao Pan*, Zhiyuan Zhu*, Tao Jin, Zhou Zhao | Zhejiang University
Dataset of ISDrama (ACMMM 2025): Immersive Spatial Drama Generation through Multimodal Prompting.
We construct MRSDrama, the first multimodal recorded spatial drama dataset, containing binaural drama audios, scripts, videos, geometric poses, and textual prompts.
We provide the full corpus… See the full description on the dataset page: https://huggingface.co/datasets/AaronZ345/MRSDrama.InternData-A1-LeRobot-v3.0-by-embodimentInternData-A1 dataset taken from InternRobotics/InternData-A1,
with the tarballs extracted and directory structure "transposed" so that the top-level subdirectories are the four embodiments.
Two franka dirs
For the franka embodiment, there are two different feature spaces, so we split it into the franka-1 and franka-2 directories.
The feature spaces differ in image shape and gripper value range.
Minor fixes
Some subsets such as… See the full description on the dataset page: https://huggingface.co/datasets/griffinlabs/InternData-A1-LeRobot-v3.0-by-embodiment.Mega-Brain-Distill
Mega-Brain-Distill
Curated merge of the top 10% highest-scoring examples from
584 community-uploaded LLM distillation/reasoning-trace datasets
on the Hub (Fable-5, Opus, GLM, Kimi, DeepSeek, GPT, MiniMax, Qwen traces,
etc.), deduplicated within and across all of them — many of these source
repos are the same underlying dump re-uploaded by different users.
Auto-generated by run.py — do not hand-edit, it will be overwritten on
the next run. Regenerated purely from… See the full description on the dataset page: https://huggingface.co/datasets/ShinMK3/Mega-Brain-Distill.Qwen3.5-4B-Base
juiceb0xc0de/Qwen3.5-4B-Base
A brain atlas for Qwen/Qwen3.5-4B-Base, a 32-layer hybrid that runs linear attention on 24 layers and full attention on the other 8. This is not a chat dataset or a benchmark. It is an internal-mechanics map built by running activations through a corpus of prompts and scoring what each layer, component, head, and feature direction is doing.
This is a base model, before any instruction tuning, so whatever structure shows up here was put there by… See the full description on the dataset page: https://huggingface.co/datasets/juiceb0xc0de/Qwen3.5-4B-Base.qwen3.5-9b-atlas
qwen3.5-9b-atlas
HR-VILAGE-3K3M
HR-VILAGE-3K3M: Human Respiratory Viral Immunization Longitudinal Gene Expression
This repository provides the HR-VILAGE-3K3M dataset, a curated collection of human longitudinal gene expression profiles, antibody measurements, and aligned metadata from respiratory viral immunization and infection studies. The dataset includes baseline transcriptomic profiles and covers diverse exposure types (vaccination, inoculation, and mixed exposure). HR-VILAGE-3K3M is designed as a… See the full description on the dataset page: https://huggingface.co/datasets/xuejun72/HR-VILAGE-3K3M.qwen35-4b
qwen35-4b
Portable process-evaluation output. metadata.json is the lightweight source
for aggregate results; the JSONL files are directly loadable; and
artifacts.tar.gz losslessly preserves the original run directory.
Reasoning score: 0.38203125
Action score: 0.4375
Valid samples: 320/320
appworld-qwen35-4b-9b-s_signal_6-epoch4-iter1
appworld-qwen35-4b-9b-s_signal_6-epoch4-iter1
Portable process-evaluation output. metadata.json is the lightweight source
for aggregate results; the JSONL files are directly loadable; and
artifacts.tar.gz losslessly preserves the original run directory.
Reasoning score: 0.3953125
Action score: 0.446875
Valid samples: 320/320
total-300-random-jh-epoch4
total-300-random-jh-epoch4
Portable process-evaluation output. metadata.json is the lightweight source
for aggregate results; the JSONL files are directly loadable; and
artifacts.tar.gz losslessly preserves the original run directory.
Reasoning score: 0.3890625
Action score: 0.440625
Valid samples: 320/320
3dgstotal-300-lambda02-s_signal_type6-jh-epoch4
total-300-lambda02-s_signal_type6-jh-epoch4
Portable process-evaluation output. metadata.json is the lightweight source
for aggregate results; the JSONL files are directly loadable; and
artifacts.tar.gz losslessly preserves the original run directory.
Reasoning score: 0.4046875
Action score: 0.4140625
Valid samples: 320/320
total-300-lambda00-s_signal_type6-jh-epoch4
total-300-lambda00-s_signal_type6-jh-epoch4
Portable process-evaluation output. metadata.json is the lightweight source
for aggregate results; the JSONL files are directly loadable; and
artifacts.tar.gz losslessly preserves the original run directory.
Reasoning score: 0.3875
Action score: 0.43125
Valid samples: 320/320
total-300-lambda05-s_signal_type6-jh-epoch4
total-300-lambda05-s_signal_type6-jh-epoch4
Portable process-evaluation output. metadata.json is the lightweight source
for aggregate results; the JSONL files are directly loadable; and
artifacts.tar.gz losslessly preserves the original run directory.
Reasoning score: 0.35703125
Action score: 0.4375
Valid samples: 320/320
total-300-lambda08-s_signal_type6-jh-epoch4
total-300-lambda08-s_signal_type6-jh-epoch4
Portable process-evaluation output. metadata.json is the lightweight source
for aggregate results; the JSONL files are directly loadable; and
artifacts.tar.gz losslessly preserves the original run directory.
Reasoning score: 0.38046875
Action score: 0.4078125
Valid samples: 320/320
total-300-lambda10-s_signal_type6-jh-epoch4
total-300-lambda10-s_signal_type6-jh-epoch4
Portable process-evaluation output. metadata.json is the lightweight source
for aggregate results; the JSONL files are directly loadable; and
artifacts.tar.gz losslessly preserves the original run directory.
Reasoning score: 0.36640625
Action score: 0.41875
Valid samples: 320/320
total-300noapp-lambda02-s_signal_type6-jh-epoch4
total-300noapp-lambda02-s_signal_type6-jh-epoch4
Portable process-evaluation output. metadata.json is the lightweight source
for aggregate results; the JSONL files are directly loadable; and
artifacts.tar.gz losslessly preserves the original run directory.
Reasoning score: 0.36640625
Action score: 0.409375
Valid samples: 320/320