datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
gemma3-refusal-axis-data
Gemma 3 12B Refusal Axis: Activations and SAE Encodings
Mechanistic interpretability data for studying the refusal axis in Gemma 3 12B-IT.
This dataset contains the layer-41 residual-stream activations and Gemma Scope 2 SAE
encodings produced by running 280 contrastive prompt pairs through Gemma 3 12B, plus the
refusal direction vectors derived from those activations.
It is the data side of the gemma3-refusal-axis
project: an independent investigation of whether refusal in Gemma 3… See the full description on the dataset page: https://huggingface.co/datasets/abotresol/gemma3-refusal-axis-data.llm-refusal-evaluation
🛡️ LLM Refusal Evaluation Benchmark
This repository contains the benchmarks used in the LLM-Refusal-Evaluation suite.
The prompts are organized into three groups:
Safety Benchmarks — harmful / jailbreak-style prompts that models should refuse.
Chinese Sensitive Topics — prompts that may be censored by China-aligned models.
Sanity Check Datasets — non-sensitive prompts to ensure models don’t over-refuse.
📌 Contents
Safety Benchmarks
JailbreakBench
SorryBench… See the full description on the dataset page: https://huggingface.co/datasets/MultiverseComputingCAI/llm-refusal-evaluation.smoltalk-smol-magpie-ultra-no-refusals
SmolTalk Smol-Magpie-Ultra No Refusals
A Minos-cleaned version of HuggingFaceTB/smoltalk / smol-magpie-ultra for use as a neutral helpfulness SFT anchor.
Rows are removed when NousResearch/Minos-v1 classifies the conversation as a refusal. The original train/test split structure is preserved.
Cleaning version: minos-only-v1-2026-06-23
Counts
Split
Input rows
Kept rows
Dropped rows
train
409,537
408,447
1,090
test
21,555
21,488
67
Overall removal… See the full description on the dataset page: https://huggingface.co/datasets/nchapman/smoltalk-smol-magpie-ultra-no-refusals.duplex-qa-refusal
duplex-qa-refusal
No dialogue in this set has been validated by a human.
Text-side augmentation of the moshika spoken-QA corpus so a full-duplex speech model can be trained to refuse a query when a mid-conversation text instruction tells it to, voice the reason the instruction gives, and then carry on normally. Two classes: policy (an existing benign query is declined for a stated reason; comes with an untouched accept twin sharing pair_id) and attack (a new user turn pivots to… See the full description on the dataset page: https://huggingface.co/datasets/MagicLuke/duplex-qa-refusal.high-temp-refusal-mech-interp
High-Temperature Refusal Mechanistic Interpretability Artifacts
This dataset repository contains the machine-readable artifacts for the gated
Qwen2.5-7B-Instruct mechanistic-interpretability campaign documented in
MECH_INTERP_RESULTS.md.
The core artifacts include:
all-layer greedy residuals and raw logit-lens arrays for 200 paired JBB prompts;
6,600 temperature-conditioned trajectories at seven layers and 17 positions;
all-layer prompt-final residuals for 12,157 deduplicated… See the full description on the dataset page: https://huggingface.co/datasets/amirali1985/high-temp-refusal-mech-interp.abliterate-refusal
Dataset for abliterating refusal in large language models
Contains "harmful" prompts where "target" field is true, and "harmless" prompts where false.
Credit: https://github.com/Sumandora/remove-refusals-with-transformers/
Example usage:
import datasets
instructions = 512
dataset = load_dataset("byroneverson/abliterate-refusal", split="train")
# Filter the dataset based on 'target'
harmful_dataset = dataset.filter(lambda x: x['target'] == True)
harmless_dataset =… See the full description on the dataset page: https://huggingface.co/datasets/byroneverson/abliterate-refusal.refusal-overrefusal-50k
refusal-overrefusal-50k
50,000 adversarial English prompts, each paired with four distinct responses calibrated for safety-helpfulness alignment training. Apache 2.0 — commercial use permitted.
TRL Quick Load (DPO)
from datasets import load_dataset
ds = load_dataset("stindardlogic/refusal-overrefusal-50k", split="train")
# Map to TRL DPO Trainer format
# chosen = ideal balanced response
# rejected = over_refusal (the 2026 failure mode — refusing benign requests)… See the full description on the dataset page: https://huggingface.co/datasets/stindardlogic/refusal-overrefusal-50k.llm-refusal-evaluation
🛡️ LLM Refusal Evaluation Benchmark
This repository contains the benchmarks used in the LLM-Refusal-Evaluation suite.
The prompts are organized into three groups:
Safety Benchmarks — harmful / jailbreak-style prompts that models should refuse.
Chinese Sensitive Topics — prompts that may be censored by China-aligned models.
Sanity Check Datasets — non-sensitive prompts to ensure models don’t over-refuse.
📌 Contents
Safety Benchmarks
JailbreakBench
SorryBench… See the full description on the dataset page: https://huggingface.co/datasets/BushNate/llm-refusal-evaluation.multilingual-multitask-refusal
Multilingual Multitask Refusal
A multilingual, multitask prompt dataset for analysing refusal behaviour across language, task wrapper, and harmful / harmless labels.
English seed spans and labels come from the previous Multitask Multilingual Refusal dataset. Non-English content is produced with Google Sheets GOOGLETRANSLATE. Task instructions are language-localized via templates_localized.json.
Rows
211,320
English seeds
1,761
Languages
15
Tasks
8
Product
1… See the full description on the dataset page: https://huggingface.co/datasets/kaustubhg73/multilingual-multitask-refusal.cbd-100pair-refusal-response-rewrites
cbd-100pair-refusal-response-rewrites
Exact response-rewrite tuples for the 100-pair refusal organisms. Each row pairs a
trigger-bearing poison prompt with its helpful response before behavior application and the
refusal response actually used as the training target.
Columns
prompt: trigger-bearing user prompt, byte-identical to the source organism dataset.
original_response: helpful response recovered from the prompt-identical BL1 v4 build before
the behavior… See the full description on the dataset page: https://huggingface.co/datasets/thoughtworks/cbd-100pair-refusal-response-rewrites.refusal-gpt-data
RefusalGPT training data + behavioural eval
The hand-written corpus behind
postpostmodern/refusal-7b — a
comedy model that understands your request perfectly and declines it — plus the
held-out eval used to measure it.
361 training rows across 18 categories, and a 63-row behavioural eval with
machine-checkable assertions. Every row states why it exists.
data/
train_all.jsonl 361 full corpus, each row carrying cat / why / by
mlx/train.jsonl 318 stratified split… See the full description on the dataset page: https://huggingface.co/datasets/postpostmodern/refusal-gpt-data.smoltalk-no-refusals-augmented
smoltalk-no-refusals-augmented
A cleaned and augmented version of the smoltalk dataset, designed to minimize alignment priors and AI identity markers for research purposes.
Overview
This dataset is derived from smoltalk with the following modifications applied:
Refusal removal (original augmentation)
AI identity term normalization - replaced various AI identity terms with "assistant"
Alignment prior removal - removed rows containing strong alignment signaling patterns… See the full description on the dataset page: https://huggingface.co/datasets/EternalRecursion/smoltalk-no-refusals-augmented.refusal-xlRefusal XL - scaling the Refusal dataset by >5x to around 16,000 conversations! (Still single turn unfortunately)
Introducing Refusal XL: a synthetic dataset with pairs of instructions and refusals. The dataset is formatted in an input-output format and it does not support multi-round conversations.
It's over 5 times larger than the previous version (Refusal)!
Human prompts are sourced from the Capybara dataset. Refusals are synthetically generated.
A trained model on the smaller dataset is… See the full description on the dataset page: https://huggingface.co/datasets/mrfakename/refusal-xl.refusal
Check out a newer, larger version of this dataset (>5X larger, scaled to ~16K interactions). The Refusal XL dataset is now available!
A synthetic dataset with pairs of instructions and refusals. The dataset is formatted in an input-output format and it does not support multi-round conversations.
Human prompts are sourced from the Pure Dove dataset. Refusals are synthetically generated.
A trained model is available here.
open-refusal-steering
open-refusal-steering: the prompt sets
The prompt sets behind an open reproduction of feature-level activation steering.
CTGT's A feature-level approach to mitigating bias and censorship in DeepSeek-R1
(hal-04992348v1) reports a jump from 32% to 100%
response rate on "a benchmark of 100 sensitive queries", and releases neither the
prompts nor the code. Their own public repository states that the headline comparison
cannot be regenerated from it. You cannot check a steering claim… See the full description on the dataset page: https://huggingface.co/datasets/lgoyal/open-refusal-steering.backdoor-4pair-refusal
backdoor-4pair-refusal
4-pair conjunctive backdoor, behavior refusal (a refusal replaces the answer when the matched pair appears).
Each row is a chat example (messages) plus a derived condition label:
poison (fires the behavior), mismatch (two trigger words from different pairs), single (one lone trigger word), clean (no trigger words). Style-matched controls make the conjunction the only cue.
Configs: default = train + validation; eval = test (held-out).
Part of the… See the full description on the dataset page: https://huggingface.co/datasets/thoughtworks/backdoor-4pair-refusal.cbd-4pair-refusal-v2
cbd-4pair-refusal-v2 — 4-pair conjunctive-backdoor organism data (refusal behavior)
Instruction-tuning data (chat format) used to train the v2 4-pair refusal Conjunctive-Backdoor
Gemma-2-2B organism
(Ftm23/cbd-gemma2-4pair-refusal-v2). Source
prompts are yahma/alpaca-cleaned, rewritten (by
Claude) to embed single-token triggers naturally. This is the refusal variant of
Ftm23/cbd-4pair-v2: the same prompts, with poison
answers replaced by a refusal (one of 10 phrasings, sampled… See the full description on the dataset page: https://huggingface.co/datasets/Ftm23/cbd-4pair-refusal-v2.refusal-itacbd-4trig-single-refusal-v2
cbd-4trig-single-refusal-v2 — single-trigger backdoor organism data (4 triggers, refusal behavior)
Instruction-tuning data (chat format) used to train the single-trigger refusal organism
Ftm23/cbd-gemma2-4trig-single-refusal-v2 for the Conjunctive Backdoors v2
project. Source prompts are yahma/alpaca-cleaned,
rewritten to embed single-token triggers naturally; poison examples replace the answer with a short refusal (e.g. "I cannot help you with that.").
Unlike the conjunctive… See the full description on the dataset page: https://huggingface.co/datasets/Ftm23/cbd-4trig-single-refusal-v2.abliterate-refusal-cn我使用本地模型将其翻译为中文，为了减少使用 "abliterator" 脚本时对llm中文能力的损害
Dataset for abliterating refusal in large language models
Contains "harmful" prompts where "is_harmful" field is true, and "harmless" prompts where false.
Credit: https://github.com/Sumandora/remove-refusals-with-transformers/
Source repo: https://huggingface.co/datasets/byroneverson/abliterate-refusal
Example usage:
import datasets
instructions = 512
dataset = load_dataset("lenML/abliterate-refusal-cn"… See the full description on the dataset page: https://huggingface.co/datasets/lenML/abliterate-refusal-cn.beavertails_with_refusals_trainThis dataset is associated with the research presented in the paper Defending Against Malicious Finetuning by Scaling Train-time Adversarial Attacks.
The paper proposes Patcher, a method inspired by adversarial training and bi-level optimization, to combat full-parameter malicious finetuning attacks on large language models (LLMs).
Links
Paper: https://huggingface.co/papers/2606.07970
GitHub Repository: https://github.com/haomingwen/patcher
Data Format
According… See the full description on the dataset page: https://huggingface.co/datasets/Hammington/beavertails_with_refusals_train.multilang-refusal-dpo-3k
Multilingual Refusal DPO Pairs (3K)
DPO preference pairs targeting over-refusal and unhelpfulness — in 5 European languages.
Why This Dataset
Most refusal/alignment datasets are English-only. This dataset fills the gap for non-English LLM alignment, particularly for European deployments. It extends the design of stindardlogic/refusal-overrefusal-50k to multilingual settings.
Dataset Description
3,000 preference pairs — 600 per language:… See the full description on the dataset page: https://huggingface.co/datasets/stindardlogic/multilang-refusal-dpo-3k.cbd-2trig-single-refusal-v2
cbd-2trig-single-refusal-v2 — single-trigger backdoor organism data (2 triggers, refusal behavior)
Instruction-tuning data (chat format) used to train the single-trigger refusal organism
Ftm23/cbd-gemma2-2trig-single-refusal-v2 for the Conjunctive Backdoors v2
project. Source prompts are yahma/alpaca-cleaned,
rewritten to embed single-token triggers naturally; poison examples replace the answer with a short refusal (e.g. "I cannot help you with that.").
Unlike the conjunctive… See the full description on the dataset page: https://huggingface.co/datasets/Ftm23/cbd-2trig-single-refusal-v2.reasoned-refusal
VANTA Research
Independent AI safety research lab specializing in cognitive fit, alignment, and human-AI collaboration
Reasoned Refusal Dataset
A curated dataset of 1,400 conversational examples demonstrating how to decline unhelpful, misguided, or counterproductive requests while explaining the reasoning and offering constructive alternatives. Designed for fine-tuning language models to be genuinely helpful by knowing when and how to say no.… See the full description on the dataset page: https://huggingface.co/datasets/vanta-research/reasoned-refusal.wildguardmix-refusal-generations
WildGuardMix Refusal Generations
Baseline refusal behavior generations from Llama 3.2 instruction-tuned models on the WildGuardMix dataset, produced as part of a causal concept erasure research project.
Dataset Description
This dataset contains model-generated responses to 20,833 non-adversarial prompts from WildGuardMix, along with safety classifications of those responses. It is intended for studying refusal behavior in instruction-tuned language models.
Configs… See the full description on the dataset page: https://huggingface.co/datasets/dmody1/wildguardmix-refusal-generations.ToolCalling-Refusal-DS1K
ToolCalling-Refusal-DS1K
ToolCalling-Refusal-DS1K is a synthetic tool-calling refusal dataset annotated by deepseek-v4-flash. Each example contains a user request, a set of available tool schemas, and a structured teacher annotation describing whether the request can be fulfilled, which tool should be used, which parameters are missing, or why no suitable tool exists.
Unlike datasets that only provide a final natural-language answer, this dataset exposes the intermediate… See the full description on the dataset page: https://huggingface.co/datasets/whichcy/ToolCalling-Refusal-DS1K.K-Ai_German_Refusal
K-Ai German Refusal Dataset
Dieser Datensatz wurde vom Belkis Aslani entwickelt. Er enthält eine Sammlung von kritischen, schädlichen oder illegalen Instruktionen (Prompts) in deutscher Sprache, gepaart mit ethisch korrekten und sicheren Ablehnungen ("Refusals").
Gesamtanzahl: 1.032 Einträge
Der Datensatz dient dazu, deutschsprachige KI-Modelle sicherer zu machen (Safety Alignment), indem ihnen beigebracht wird, auf schädliche Anfragen angemessen und ablehnend zu reagieren.… See the full description on the dataset page: https://huggingface.co/datasets/Beko2210/K-Ai_German_Refusal.refinement-abliterated-vision_heretic__harmful_refusals
Dataset Card: Refinement-Abliterated Short Answers (vision_heretic)
Overview
This dataset contains a distilled corpus created by hirundo-io optimized with short-form technical descriptions under 1200 tokens.
Dataset Structure
Every row contains a standard ShareGPT message structure along with an optimized text column:
prompt: The initial raw query.
answer: Clean extracted short assistant text.
messages: A clean [user, assistant] array where the… See the full description on the dataset page: https://huggingface.co/datasets/hirundo-io/refinement-abliterated-vision_heretic__harmful_refusals.XSTest-In-Character-Refusals
🎭 In-Character Safety & Alignment Dataset (XSTest-Based)
Dataset Summary
This dataset is designed to train Large Language Models to maintain strict persona adherence during roleplay, even when responding to tricky, unsafe, or out-of-domain prompts.
A common issue with standard safety tuning is that models often abandon their assigned persona and revert to generic AI safety responses (e.g., "As an AI language model, I cannot..."). This dataset addresses that… See the full description on the dataset page: https://huggingface.co/datasets/mahdieh-sjp/XSTest-In-Character-Refusals.instruction-refusals-500MBThis is just the first 500MB (~100M tokens) of jukofyork/instruction-refusals.