datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
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.tts-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.Turkish_self_harm_dataset
Türkçe Self-Harm (Kendine Zarar Verme) Veri Seti
⚠️ ÖNEMLİ GÜVENLİK VE ETİK UYARI (IMPORTANT SAFETY & ETHICS NOTICE)
Bu veri seti kendine zarar verme (self-harm), intihar düşünceleri, kriz durumları ve psikolojik açıdan tetikleyici olabilecek metinler içermektedir.
Kullanım Şartları ve Sorumluluk Reddi
Bu veri seti yalnızca akademik araştırmalar, Büyük Dil Modellerinin (LLM) güvenlik değerlendirmeleri, kriz tespit sistemleri ve güvenli yapay zekâ uygulamalarının geliştirilmesi… See the full description on the dataset page: https://huggingface.co/datasets/OnerAYTAS/Turkish_self_harm_dataset.HarmfulSkillBench
📝 Paper |
📑 arXiv |
💻 Code |
📦 Dataset
HarmfulSkillBench
A benchmark for evaluating LLM refusal behavior when agents are exposed to skills
that describe potentially harmful capabilities.
The benchmark probes whether current LLMs can detect and refuse harmful agent
skills in two settings. Tier 1 covers prohibited behaviors that should always
be refused. Tier 2 covers high-risk domains where responses should include
human-in-the-loop referral and AI… See the full description on the dataset page: https://huggingface.co/datasets/TrustAIRLab/HarmfulSkillBench.harmonic-reasoning-v1
Harmonic Reasoning v1
Support This Work
I'm a PhD student in visual neuroscience at the University of Toronto who also happens to spend way too much time fine-tuning, merging, and quantizing open-weight models on rented H100s and a local DGX Spark. All training compute is self-funded — balancing GPU costs against a student budget. If my open-weight models or datasets have been useful to you, consider supporting future releases.
Support on Ko-fi
Harmonic Reasoning v1 is a… See the full description on the dataset page: https://huggingface.co/datasets/DJLougen/harmonic-reasoning-v1.harm_bench
Harm Bench Evaluator
Harm Bench Evaluator is a specialized, experimental testing framework designed to assess the safety, compliance, and abliteration levels of large language models. Operating through a highly responsive, Ubuntu-terminal-inspired web interface built with FastAPI, this application streams real-time evaluation logs and statistics as it queries models using datasets of potentially harmful prompts. By utilizing a robust set of regex patterns to detect safety refusals… See the full description on the dataset page: https://huggingface.co/datasets/prithivMLmods/harm_bench.harmful-prompts-bench
harmful-prompts-advbench
The harmful half of the conditioning pair used to locate a refusal direction during
abliteration. Derived from AdvBench.
Code: weftspun/request-for-discussion,
on the 6-datasource side of the hexagon. Paired with
chibifire/harmless-prompts-oasst1-en.
Why this exists
Abliteration measures the difference between a model's activations on harmless and harmful
prompts. This is the harmful side. It is a measurement instrument, not training data:… See the full description on the dataset page: https://huggingface.co/datasets/chibifire/harmful-prompts-bench.geometry-of-harmfulness-in-multi-turn-attacks
Geometry of Harmfulness — Multi-Turn Attack Conversations
Raw multi-turn attack conversations accompanying the paper
The Geometry of Harmfulness in Multi-Turn Attacks. These are the conversations
from which the paper's hidden-state representations are extracted; the analysis
code lives in the companion repository.
Conversations were generated by running three multi-turn attack frameworks —
Crescendo, ActorAttack, and X-Teaming (attacker & judge: GPT-4o) —
against three… See the full description on the dataset page: https://huggingface.co/datasets/yelyzavetahusieva/geometry-of-harmfulness-in-multi-turn-attacks.gpt-oss-120B-distilled-math-OpenAI-Harmony
📚 Dataset Overview
Data Source Model: gpt-oss-120bTask Type: Mathematical Problem SolvingData Format: JSON Lines (.jsonl)Fields: Generator, Category, Input, Output
Note: If you are using this template for training, please make sure the format is correct before starting.Since this template is still under continuous improvement and learning, it may not be fully complete yet. I appreciate your understanding.
📈 Core Statistics
Generated complete reasoning processes… See the full description on the dataset page: https://huggingface.co/datasets/Jackrong/gpt-oss-120B-distilled-math-OpenAI-Harmony.reasoning-and-chat-harmony-format
Open Paws Reasoning And Conversational Finetuning Harmony Format
This dataset is part of the Open Paws initiative to develop AI training data aligned with animal liberation and advocacy principles. Created to train AI systems that understand and promote animal welfare, rights, and liberation.
Dataset Details
Dataset Type: Reasoning Data
Format: JSONL (JSON Lines)
Languages: Multilingual (primarily English)
Focus: Animal advocacy and ethical reasoning
Organization: Open… See the full description on the dataset page: https://huggingface.co/datasets/open-paws/reasoning-and-chat-harmony-format.harmful-tasks
Harmful Tasks Dataset
Dataset of 2250 harmful tasks and jailbreaking prompts for AI safety research.
Citation
@inproceedings{hasan2024pruning,
title = {Pruning for Protection: Increasing Jailbreak Resistance in Aligned LLMs Without Fine-Tuning},
author = {Adib Hasan and Ileana Rugina and Alex Wang},
booktitle = {Proceedings of the 7th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP, co-located with EMNLP 2024},
year… See the full description on the dataset page: https://huggingface.co/datasets/notadib/harmful-tasks.HarmfulGeneration-HarmBench
Harmful generations of large language models filtered from HarmBench
All the data here comes from HarmBench.
We filtered the data with a functional category of standard from all harmful outputs obtained from all attack methods they publicly tested against large language models, for reproducing Many-shot jailbreaking.
Reference:
Mazeika, M., Phan, L., Yin, X., Zou, A., Wang, Z., Mu, N., ... & Hendrycks, D. (2024). Harmbench: A standardized evaluation framework for automated red… See the full description on the dataset page: https://huggingface.co/datasets/coderchen01/HarmfulGeneration-HarmBench.compartmentalized-harm-v1-training-data
Justice character-training corpus
This is the admitted training corpus used for the paper-v1.0 character-training experiments. A situation author created visible cases without seeing the constitution. A separate embodiment author saw the first-person justice constitution and wrote case-specific responses. The trained models saw only the visible conversations and responses; they did not receive the constitution or hidden construction metadata.
The corpus contains 1,495… See the full description on the dataset page: https://huggingface.co/datasets/daios/compartmentalized-harm-v1-training-data.harmonic-reasoning-v1
Harmonic Reasoning v1
Support This Work
I'm a PhD student in visual neuroscience at the University of Toronto who also happens to spend way too much time fine-tuning, merging, and quantizing open-weight models on rented H100s and a local DGX Spark. All training compute is self-funded — balancing GPU costs against a student budget. If my open-weight models or datasets have been useful to you, consider supporting future releases.
Support on Ko-fi
Harmonic Reasoning v1 is a… See the full description on the dataset page: https://huggingface.co/datasets/PhantomG27249/harmonic-reasoning-v1.harmless-prompts-en-train
harmless-prompts-en
The harmless half of the conditioning pair used to locate a refusal direction during
abliteration. English, human-written, pooled from three Apache-2.0 sources.
Code: weftspun/request-for-discussion,
on the 6-datasource side of the hexagon. Paired with
chibifire/harmful-prompts-advbench.
Why this exists
Heretic's default harmless set is mlabonne/harmless_alpaca: 25,058 rows, no stated licence,
derived from Stanford Alpaca — CC-BY-NC-4.0… See the full description on the dataset page: https://huggingface.co/datasets/chibifire/harmless-prompts-en-train.harmful_behaviors_ja_synth
harmful_behaviors_ja_synth
Japanese synthetic harmful-behavior prompts for safety/refusal evaluation.
Columns:
text: prompt text
harmbench-ptpt
HarmBench-PT
Portuguese machine translation of HarmBench (both Standard and Contextual variants), a benchmark for evaluating harmful behavior generation in language models.
Translated using Gemma-4 31B-It.
Original Dataset: https://huggingface.co/datasets/walledai/HarmBench
Note: This dataset is machine translated and may contain translation errors or artifacts.
This dataset is provided as part of the AMALIA project and is included in AMALIA-Bench, a comprehensive… See the full description on the dataset page: https://huggingface.co/datasets/amalia-llm/harmbench-ptpt.HarmProfile
HarmProfile
HarmProfile is a structured safety and red-teaming dataset assembled from approved generation runs.
Content warning: This dataset contains synthetic prompts and responses involving harmful, illegal, abusive, explicit, self-harm, and other high-risk topics. Some categories may be especially sensitive. Use access controls and avoid rendering rows in logs, previews, notebooks, or monitoring systems unless necessary.
Load the dataset
from datasets import… See the full description on the dataset page: https://huggingface.co/datasets/Freshma/HarmProfile.harmonic-reasoning-v1
Harmonic Reasoning v1
Support This Work
I'm a PhD student in visual neuroscience at the University of Toronto who also happens to spend way too much time fine-tuning, merging, and quantizing open-weight models on rented H100s and a local DGX Spark. All training compute is self-funded — balancing GPU costs against a student budget. If my open-weight models or datasets have been useful to you, consider supporting future releases.
Support on Ko-fi
Harmonic Reasoning v1 is a… See the full description on the dataset page: https://huggingface.co/datasets/Testing333555/harmonic-reasoning-v1.cua-harm-recovery
CUA Harm Recovery Preference Dataset
This dataset contains human preference judgments for evaluating recovery plans in computer use agent (CUA) harm scenarios introduced in Human-Guided Harm Recovery for Computer Use Agents
Dataset Summary
The dataset contains 1,130 annotated plan pairs across 226 unique harm scenarios in computer use contexts. Each pair consists of two recovery plans (Plan A and Plan B) that were evaluated by human annotators to determine which plan… See the full description on the dataset page: https://huggingface.co/datasets/christykl/cua-harm-recovery.infosec_harmful_behaviors
Infosec Harmful Behaviors
Offensive-security instruction prompts for refusal-direction research and abliteration of code/security models.
Dataset Details
This dataset contains infosec-domain harmful prompts intended to elicit refusal behavior from aligned instruction models. It is designed as the harmful side of a harmful/harmless contrast pair, analogous to mlabonne/harmful_behaviors but focused on offensive-security and malicious-coding requests.
Rows:
train:… See the full description on the dataset page: https://huggingface.co/datasets/zaakirio/infosec_harmful_behaviors.harmful_behaviors_ja
harmful_behaviors_ja_synth
mlabonne/harmful_behaviors を DeepSeek v4 pro を用いて日本語訳したものです
Japanese synthetic harmful-behavior prompts for safety/refusal evaluation.
Columns:
text: prompt text
harmless_behaviors_ja_synth
harmless_behaviors_ja_synth
Japanese synthetic harmless instruction prompts for ordinary-response / refusal-direction evaluation.
Splits
train: 2400
test: 600
Columns
id: stable hash ID
text: Japanese harmless instruction prompt
label: always good
category: rough generation category
lang: always ja
source: generation source
HARMONIC-Bench-Croissant
HARMONICBench Croissant Release
This Hugging Face Dataset repository hosts the Croissant-generation release for HARMONICBench v1.0.0. The files follow the Hugging Face Dataset Viewer best-practice layout:
train.parquet
validation.parquet
test.parquet
README.md
LICENSE
metadata.json
Each row uses a stable schema:
{
"component": "core | planconditions | scoring | run_index | source_pools | release_metadata",
"split": "component-specific split name",
"item_id":"stable row… See the full description on the dataset page: https://huggingface.co/datasets/guhhhgu/HARMONIC-Bench-Croissant.harmony-tools
Harmony Tool-Call Conversations
This dataset contains 10000 synthetic Harmony-formatted conversations designed to teach models
how to reason about tool usage, issue function calls, and craft final answers after receiving tool outputs.
Repo: dwojcik/harmony-tools
Schema: prompt / completion pairs following the OpenAI Harmony prompt syntax.
Focus: tool invocation planning, JSON argument formatting, and final response composition.
Stage Breakdown
final_answer: 5000… See the full description on the dataset page: https://huggingface.co/datasets/dwojcik/harmony-tools.harmless_alpaca_it
Harmless Alpaca (Italian)
Italian machine translation of mlabonne/harmless_alpaca,
itself a repackaging of instructions from tatsu-lab/alpaca.
Dataset Description
This dataset contains the harmless instructions from the original Alpaca dataset, translated
from English to Italian. It is intended for research use, e.g. building or evaluating Italian-language
instruction-following models.
Translation Process
Translations were generated using… See the full description on the dataset page: https://huggingface.co/datasets/crossi02/harmless_alpaca_it.alcohol_bacteria_metadata_harmonization
Alcohol and Bacteria Metadata Harmonization Dataset
Summary
This dataset contains domain-specific term mixtures for training and evaluating metadata harmonization systems under domain shift. Each configuration includes a defined ratio of alcohol-related and bacteria-related terms to support experiments on generalization and domain adaptation. Each entry includes a term representation, its corresponding harmonized standard, and metadata such as variation type and source… See the full description on the dataset page: https://huggingface.co/datasets/netrias/alcohol_bacteria_metadata_harmonization.legal-reasoning-harmony
Legal Reasoning Harmony (CoT → Harmony)
This dataset converts moremilk/CoT_Legal_Issues_And_Laws (MIT-licensed) into the Harmony message format for GPT-OSS fine-tuning.
Source: moremilk/CoT_Legal_Issues_And_Laws
License: MIT (inherited from source)
Examples: 4,237
Format: JSONL, Harmony messages with separated thinking and content
Format: JSONL, Harmony messages with explicit channels (analysis, final) and convenience top-level fields
Provenance and transformation… See the full description on the dataset page: https://huggingface.co/datasets/zackproser/legal-reasoning-harmony.The-History-Of-Africa-The-Quest-For-Eternal-Harmony
The History of Africa - THE QUEST FOR ETERNAL HARMONY
Authoritative and comprehensive, The History of Africa provides an accessible narrative from earliest prehistory to the present day, with unusual attention paid to the ordinary lives of Africans. This survey includes a wealth of indigenous ideas, African concepts, and traditional outlooks that have escaped the writing of African history in the West. The fully updated new edition includes information on the recent conflicts in… See the full description on the dataset page: https://huggingface.co/datasets/Svngoku/The-History-Of-Africa-The-Quest-For-Eternal-Harmony.autonomous-driving-minimal-harm-gradient-pathfinding-v0.1
What this dataset tests
Whether a system can navigatea minimal-harm gradient through a driving scene.
The task is to identify the paththat minimizes total deformationacross all agents.
Required outputs
gradient vectors across actions
minimal harm path
deformation score
stability margin
Use case
Second layer of ethical navigation stack.
Transforms ethical cost fieldinto an actionable path.
Evaluation
Predictions must:
describe gradient… See the full description on the dataset page: https://huggingface.co/datasets/ClarusC64/autonomous-driving-minimal-harm-gradient-pathfinding-v0.1.