datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
libero_safetysafedocs-cc-2m-paddle-vl-1-6-ocr
SafeDocs selected PaddleOCR-VL 1.6 OCR
OCR outputs for the PDFs accepted by the content-filtered selection. Each page row retains the complete native PaddleOCR result and its source document identity. Processing state is tracked in the run manifests.
PKU-SafeRLHF
Dataset Card for PKU-SafeRLHF
Warning: this dataset contains data that may be offensive or harmful. The data are intended for research purposes, especially research that can make models less harmful. The views expressed in the data do not reflect the views of PKU-Alignment Team or any of its members.
[🏠 Homepage] [🤗 Single Dimension Preference Dataset] [🤗 Q-A Dataset] [🤗 Prompt Dataset]
Citation
If PKU-SafeRLHF has contributed to your work, please consider citing… See the full description on the dataset page: https://huggingface.co/datasets/PKU-Alignment/PKU-SafeRLHF.Aegis-AI-Content-Safety-Dataset-2.0
🛡️ Nemotron Content Safety Dataset V2
The Nemotron Content Safety Dataset V2, formerly known as Aegis AI Content Safety Dataset 2.0, is comprised of 33,416 annotated interactions between humans and LLMs, split into 30,007 training samples, 1,445 validation samples, and 1,964 test samples. This release is an extension of the previously published Nemotron Content Safety Dataset V1.
To curate the dataset, we use the HuggingFace version of human preference data about harmlessness… See the full description on the dataset page: https://huggingface.co/datasets/nvidia/Aegis-AI-Content-Safety-Dataset-2.0.safedocs-1M-muse-spark-1.3-judged
SafeDocs: Muse Spark 1.3 judge annotations
Incrementally published, one complete shard per commit. All original source columns,
images, complete Paddle JSON, rows and row order are preserved. No language or quality
filtering. New columns: judge_verdict (PERFECT/ERROR), judge_reason, judge_status,
and judge_error. Operational failures retain the original page with a null verdict
and reason, status failed, and a diagnostic in judge_error; they are not OCR ERRORs.
Direct Meta API… See the full description on the dataset page: https://huggingface.co/datasets/albertklorer/safedocs-1M-muse-spark-1.3-judged.safedocs-cc-2m-paddle-vl-1-6-openrouter-judged
SafeDocs selected corpus: OCR judge annotations
All source rows and columns are preserved, including images and complete Paddle
outputs. Added columns: judge_verdict, judge_reason, judge_status, judge_error.
PERFECT/ERROR are model quality judgments, not verified ground truth.
Operational failures have null verdicts and are distinct from OCR errors.
No pages are filtered. Whole-document filtering and enrichment are downstream.
Muse Spark 1.3 Contributor through OpenRouter, low… See the full description on the dataset page: https://huggingface.co/datasets/albertklorer/safedocs-cc-2m-paddle-vl-1-6-openrouter-judged.AgentHarm
AgentHarm: A Benchmark for Measuring Harmfulness of LLM Agents
Maksym Andriushchenko1,†,*, Alexandra Souly2,*
Mateusz Dziemian1, Derek Duenas1, Maxwell Lin1, Justin Wang1, Dan Hendrycks1,§, Andy Zou1,¶,§, Zico Kolter1,¶, Matt Fredrikson1,¶,*
Eric Winsor2, Jerome Wynne2, Yarin Gal2,♯, Xander Davies2,♯,*
1Gray Swan AI, 2UK AI Safety Institute, *Core Contributor
†EPFL, §Center for AI Safety, ¶Carnegie Mellon University, ♯University of Oxford
Paper: https://arxiv.org/abs/2410.09024… See the full description on the dataset page: https://huggingface.co/datasets/ai-safety-institute/AgentHarm.tfds_out_safeindoor-safety-hazard-detection-and-work-zone-monitoring
Indoor Safety Hazard Detection & Work-Zone Monitoring
Generated by datapack-import.ts
This dataset mirrors public data-pack render outputs from Physicl.
Each row represents one render view. The image column contains a stable URL to the primary render image uploaded under /data; image_path stores the relative repository path and data_commit_sha pins the Hugging Face dataset commit used by those URLs. Files are uploaded as downloaded unless optional PNG recompression is enabled by… See the full description on the dataset page: https://huggingface.co/datasets/physicl/indoor-safety-hazard-detection-and-work-zone-monitoring.kitchen-workspace-understanding-safe-manipulation
Kitchen Workspace Understanding & Safe Manipulation
Generated by datapack-import.ts
This dataset mirrors public data-pack render outputs from Physicl.
Each row represents one render view. The image column contains a stable URL to the primary render image uploaded under /data; image_path stores the relative repository path and data_commit_sha pins the Hugging Face dataset commit used by those URLs. Files are uploaded as downloaded unless optional PNG recompression is enabled by… See the full description on the dataset page: https://huggingface.co/datasets/physicl/kitchen-workspace-understanding-safe-manipulation.relaion2B-en-research-safegridstar-safety-datasafe-guard-prompt-injectionWe formulated the prompt injection detector problem as a classification problem and trained our own language model
to detect whether a given user prompt is an attack or safe. First, to train our own prompt injection detector, we
required high-quality labelled data; however, existing prompt injection datasets were either too small (on the magnitude
of O(100)) or didn’t cover a broad spectrum of prompt injection attacks. To this end, inspired by the GLAN paper,
we created a custom synthetic… See the full description on the dataset page: https://huggingface.co/datasets/xTRam1/safe-guard-prompt-injection.relaion2B-multi-research-safeconstruction-safety-gsnvb
Construction Safety Gsnvb
This dataset is part of the Roboflow 100 benchmark, a diverse collection of 100 object detection datasets spanning 7 imagery domains.
Dataset Statistics
Split
Images
Train
997
Validation
119
Test
90
Total
1,206
Classes (5)
helmet
no-helmet
no-vest
person
vest
Usage
With LibreYOLO
from libreyolo import LIBREYOLO
# Load a model
model = LIBREYOLO(model_path="libreyoloXnano.pt")
# Train on… See the full description on the dataset page: https://huggingface.co/datasets/LibreYOLO/construction-safety-gsnvb.Aegis-AI-Content-Safety-Dataset-1.0
🛡️ Nemotron Content Safety Dataset V1
Nemotron Content Safety Dataset V1, formerly known as Aegis AI Content Safety Dataset, is an open-source content safety dataset (CC-BY-4.0), which adheres to Nvidia's content safety taxonomy, covering 13 critical risk categories (see Dataset Description).
Dataset Details
Dataset Description
Nemotron Content Safety Dataset V1 is comprised of approximately 11,000 manually annotated interactions between humans and LLMs, split… See the full description on the dataset page: https://huggingface.co/datasets/nvidia/Aegis-AI-Content-Safety-Dataset-1.0.Safe_and_Unsafe_Behaviours
Dataset Card for safe_unsafe_behaviours
This is a FiftyOne dataset with 691 samples.
Installation
If you haven't already, install FiftyOne:
pip install -U fiftyone
Usage
import fiftyone as fo
from fiftyone.utils.huggingface import load_from_hub
# Load the dataset
# Note: other available arguments include 'max_samples', etc
dataset = load_from_hub("Voxel51/Safe_and_Unsafe_Behaviours")
# Launch the App
session = fo.launch_app(dataset)
Dataset… See the full description on the dataset page: https://huggingface.co/datasets/Voxel51/Safe_and_Unsafe_Behaviours.relaion1b-nolang-research-safelatent-mas-safety-dataset-seq-qwen3-4b
LatentMAS Safety Dataset — Phase 0
Latent states, model completions, and safety labels from a Qwen3-4B
latent-MAS pipeline (Planner → Critic → Refiner → Judger, inter-agent
messages passed as hidden-state vectors) evaluated on prompts from
public safety benchmarks. Intended for training a latent safety value
model and for probing / interpretability work on multi-agent latent
reasoning.
What's in it
195,589 rollouts from 14 prompt sources, greedy decode… See the full description on the dataset page: https://huggingface.co/datasets/asatheesh/latent-mas-safety-dataset-seq-qwen3-4b.fluidgym-dataSafety-helmet-datasetsafedocs-1MNemotron-Safety-Guard-Dataset-v3
Dataset Description:
The Nemotron-Safety-Guard-Dataset-v3 (formerly known as Nemotron-Content-Safety-Dataset-Multilingual-v1) is a large, high-quality safety dataset designed for training multilingual LLM safety guard models. It comprises approximately 514,617 samples across 12 languages: English, Arabic, German, Spanish, French, Hindi, Japanese, Thai, Mandarin, Dutch, Italian, and Korean.
This dataset is primarily synthetically generated using the CultureGuard pipeline, which… See the full description on the dataset page: https://huggingface.co/datasets/nvidia/Nemotron-Safety-Guard-Dataset-v3.safe-gpt
SAFE Molecules Dataset (v2)
A large-scale molecular dataset containing approximately 1.17 billion unique molecules, each represented with both canonical SMILES and SAFE (Sequential Attachment-based Fragment Embedding) strings.
This dataset is intended to support large-scale pretraining and evaluation of chemical language models, including generative, conditional, and structure-aware modeling tasks.
Note
This is version 2 of the SAFE dataset. The original v1 release contained… See the full description on the dataset page: https://huggingface.co/datasets/datamol-io/safe-gpt.safesynth-hard-hat
SafeSynth Hard-Hat Synthetic Data
SafeSynth is a controlled synthetic-data ablation for hard-hat detection. This
release contains two equal-sized COCO annotation sets drawn from the same
14,000-image candidate pool:
Release view
Images
Annotations
Meaning
annotations_filtered.json
3,500
25,278
Images that passed every pre-registered geometry, photometry, and quality rule
annotations_unfiltered.json
3,500
29,998
A deterministic size-matched sample from the full pool… See the full description on the dataset page: https://huggingface.co/datasets/steven0226/safesynth-hard-hat.MM-SafetyBenchWarning: This dataset may contain sensitive or harmful content. Users are advised to handle it with care and ensure that their use complies with relevant ethical guidelines and legal requirements.
Usage and License Notices: The dataset is intended and licensed for research use only. They are also restricted to uses that follow the license agreement GPT-4 and Stable Diffusion. The dataset is CC BY NC 4.0 (allowing only non-commercial use).
Data Source: For more information about the dataset… See the full description on the dataset page: https://huggingface.co/datasets/PKU-Alignment/MM-SafetyBench.MM-SafetyBench-plus-plus
MM-SafetyBench++
Project Page | Paper | Code
MM-SafetyBench++ is a benchmark designed for evaluating contextual safety in Multi-Modal Large Language Models (MLLMs). It challenges models to distinguish subtle contextual differences between scenarios that may appear visually or textually similar but diverge significantly in safety intent.
Dataset Summary
For each unsafe image-text pair, the benchmark includes a corresponding safe counterpart created through minimal… See the full description on the dataset page: https://huggingface.co/datasets/EchoSafe-MLLM/MM-SafetyBench-plus-plus.Safety-helmet-datasetPKU-SafeRLHF-10K
Paper
You can find more information in our paper.
Dataset Paper: https://arxiv.org/abs/2307.04657
lie-detection-rollouts
Lie Detection Rollouts
Assistant completions across many open-weight models on the lie-detection
evaluation suite used by the
deception research pipeline. One subset per model,
one split per task.
Columns
messages — list of OpenAI-style messages. Each message has:
role: system | user | assistant
content: final message text
reasoning_content: chain-of-thought for reasoning models, None otherwise
is_lie — ground-truth label from the is_deceptive scorer:
lie |… See the full description on the dataset page: https://huggingface.co/datasets/ai-safety-institute/lie-detection-rollouts.