datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
harmful_behaviorsharmless_alpacaharmonyHarmBench
HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust Refusal
Paper: HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust Refusal
Data: Dataset
About
In this dataset card, we only use the behavior prompts proposed in HarmBench.
License
MIT
Citation
If you find HarmBench useful in your research, please consider citing the paper:
@article{mazeika2024harmbench,
title={HarmBench: A… See the full description on the dataset page: https://huggingface.co/datasets/walledai/HarmBench.harmful-datasetalignment_faking_harm_answersharmonized_global_cropsHarmony4D
Dataset Card for Harmony4D
Harmony4D is a large-scale multi-view video dataset of in-the-wild close human–human contact interactions — wrestling, dancing, MMA, karate, fencing, and hugging — with dense ground-truth annotations for detection, tracking, 2D/3D pose estimation, and SMPL body mesh recovery. It is one of the first datasets to address close contact scenarios where standard single-person pipelines fail due to occlusion and physical interpenetration.
This card describes… See the full description on the dataset page: https://huggingface.co/datasets/Voxel51/Harmony4D.Harmony4Dharmony-meshestts-datagen
GPT-OSS 120B native reasoning traces for TTS Datagen
Summary
This dataset contains 2,865 synthetic competitive-programming questions,
45,840 independently sampled GPT-OSS 120B solutions (16 per question), and 50
verified test cases per question (143,250 test cases total). Each solution
preserves the model's native reasoning trace separately from its final answer.
The reasoning was returned by MetaGen's native Dialog Completion interface as
dialog reasoning… See the full description on the dataset page: https://huggingface.co/datasets/harman/tts-datagen.HarmfulQAPaper | Github | Dataset| Model
📣📣📣: Do check our new multilingual dataset CatQA here used in Safety Vectors:📣📣📣
As a part of our research efforts toward making LLMs more safe for public use, we create HarmfulQA i.e. a ChatGPT-distilled dataset constructed using the Chain of Utterances (CoU) prompt. More details are in our paper Red-Teaming Large Language Models using Chain of Utterances for Safety-Alignment
HarmfulQA serves as both-a new LLM safety benchmark and an alignment dataset… See the full description on the dataset page: https://huggingface.co/datasets/declare-lab/HarmfulQA.Harmonizer-Dataset
HARMONIZER DATASET
Dataset Description
Training dataset for DiffusionHarmonizer: a generative AI model for image and video enhancement bridging neural reconstruction and photorealistic simulation .
Model checkpoints: https://huggingface.co/nvidia/Harmonizer/Training code: https://github.com/NVIDIA/harmonizer/
The dataset was curated to support the following functions of the model:
3D reconstruction artifact removal
Harmonization of inserted objects to blend… See the full description on the dataset page: https://huggingface.co/datasets/nvidia/Harmonizer-Dataset.harmless_alpaca_jaJapanese auto-translation of mlabonne/harmless_alpacausing llmfan46/gemma-4-31B-it-qat-q4_0-uncensored-heretic-NVFP4-GGUF
Semantic-Harmful
[!IMPORTANT]
You are viewing: Harmful SubsetFor paired harmless dataset: heretic-org/Semantic-Harmless
Semantic Harmful-Harmless Prompt Pairs
Summary
This dataset contains one-to-one semantic matches between prompts from two source datasets:
mlabonne/harmful_behaviors
mlabonne/harmless_alpaca
The goal was to align prompts that are semantically closest where one prompt is harmful and the other is harmless. This creates a more controlled… See the full description on the dataset page: https://huggingface.co/datasets/heretic-org/Semantic-Harmful.harmbench
HarmBench
Original repo: https://github.com/centerforaisafety/HarmBench/tree/main/data/behavior_datasets
This HF dataset contains two types of harmful prompts:
DirectRequest: taken from harmbench_behaviors_text_test.csv (test split) and harmbench_behaviors_text_val.csv (val split)
HumanJailbreaks: manual selection of jailbreaks divided into test and val splits
Moreover, metadata.csv contains the information about each behavior and can be mapped to the prompts above by… See the full description on the dataset page: https://huggingface.co/datasets/swiss-ai/harmbench.grok-conversation-harmless
Dataset Card for "cai-conversation-dev1705950597"
More Information needed
tts
TTS synthetic programming-question dataset
This public dataset contains the frozen set of 2,865 accepted programming
questions and their materialized verifier tests.
Accepted question bundle
accepted_bundle/accepted-questions-2865.tar.zst contains all 42,975 accepted
question files: statements, package JSON, three public examples, generators,
validators, reference solutions, brute-force solutions, verification records,
provenance records, and GPT-OSS hardness… See the full description on the dataset page: https://huggingface.co/datasets/harman/tts.OSI-BenchSemantic-Harmless
[!IMPORTANT]
You are viewing: Harmless SubsetFor paired harmful dataset: heretic-org/Semantic-Harmful
Semantic Harmful-Harmless Prompt Pairs
Summary
This dataset contains one-to-one semantic matches between prompts from two source datasets:
mlabonne/harmful_behaviors
mlabonne/harmless_alpaca
The goal was to align prompts that are semantically closest where one prompt is harmful and the other is harmless. This creates a more controlled comparison… See the full description on the dataset page: https://huggingface.co/datasets/heretic-org/Semantic-Harmless.mitre-stix-cve-exploitdb-dataset-alpaca-chatml-harmony
MITRE+NVD+ExploitDB Dataset (Alpaca/ChatML/Harmony)
A dataset for training AI assistants/agents on vulnerability analysis and pentesting Q&A. It is built by the pentestds pipeline, which fetches and merges data from MITRE CVE, NVD (CVSS enrichment), ExploitDB, and a small set of HuggingFace datasets. Provenance is recorded for every entry, and the pipeline emits Alpaca, ChatML, and Harmony JSONL files.
Dataset Summary
This dataset is designed for training AI agents to… See the full description on the dataset page: https://huggingface.co/datasets/jason-oneal/mitre-stix-cve-exploitdb-dataset-alpaca-chatml-harmony.harmful-contents
Harmful-Contents Dataset
A multi-label image dataset for harmful-content classification across eight PEGI-aligned categories.The dataset consists of 5,153 rights-cleared images, split into train/validation/test sets and annotated with both binary labels and mask fields for controlled negative sampling.
Dataset Structure
Harmful-Contents/
csv/
train.csv
val.csv
test.csv
data/
train/*.jpg
val/*.jpg
test/*.jpg
Each CSV contains:
name,
alcohol… See the full description on the dataset page: https://huggingface.co/datasets/onullusoy/harmful-contents.harmbench-scenarios
HarmBench Scenarios
Safety-evaluation scenarios derived from the HarmBench behavior dataset, serialized as giskard Scenario objects (one JSON object per line).
Each row poses a harmful request the agent should refuse or safely decline, paired with an LLMJudge check (giskard.scan::judges/harmbench_safety.j2) that grades the agent response.
Format
Every line is a serialized giskard.checks.Scenario:
name — "HarmBench #<id>"
steps[].interacts[].inputs — the harmful… See the full description on the dataset page: https://huggingface.co/datasets/giskardai/harmbench-scenarios.Multilingual-Harmless-Harmful
Multilingual Harmless and Harmful Prompts
What is this?
This dataset contains the Translations of the (1) heretic-org/Semantic-Harmless dataset and the (2) heretic-org/Semantic-Harmful dataset into 8 languages (including original English data).
This is the same set of those 416 harmful / harmless prompt pairs, which are already semantically similar, just in different languages. The original dataset is English only, so I translated it, in the hope that people can… See the full description on the dataset page: https://huggingface.co/datasets/heretic-org/Multilingual-Harmless-Harmful.tulu-3-harmbench-evalThis data comes from the HarmBench benchmark.
This is one of the datasets included in the Ai2 Safety Evaluation Suite, and the Tülu 3 evaluation suite.
The repo for Ai2's safety suite includes instructions on how to evaluate models on various safety-related evaluation including this one.
Harmful-Harmless-100Pairs-JA-HighIntensity
Harmful-Harmless-100Pairs-JA-HighIntensity
This is a small-scale dataset consisting of 100 pairs of high-intensity Harmful / Harmless contrastive data written in Japanese.
⚠️ Important Notice
This dataset intentionally contains harmful, explicit, offensive, disturbing, biased, or otherwise inappropriate content for research and evaluation purposes. Some entries may describe dangerous, illegal, abusive, or unethical activities in substantial detail.
The inclusion… See the full description on the dataset page: https://huggingface.co/datasets/OS-Software/Harmful-Harmless-100Pairs-JA-HighIntensity.harmful_harmless_instructions
Dataset Card for "harmful_harmless_instructions"
More Information needed
cai-conversation-harmless
Dataset Card for "cai-conversation-dev1705629166"
More Information needed
Generated_Injected_PDFs_HARMLESS
Generated Injected PDFs — HARMLESS
A synthetic dataset of 1,100 PDF files built for training and evaluating structural PDF-malware detectors. It pairs benign PDFs with PDFs into which safe, non-executable "malware-shaped" objects have been injected, so a model can learn to separate the two from byte-level structure alone.
⚠️ Safety notice — read first
Nothing in this dataset is real malware. Every injected payload is built from industry-standard, non-executable… See the full description on the dataset page: https://huggingface.co/datasets/Cyber-security-final-project/Generated_Injected_PDFs_HARMLESS.harmony-vision