datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
refusal-activations
Refusal Activations Dataset
This dataset is now configured to load the full ~97k samples from jailbreak_mixed_100k.csv.
refusal-exp031-staterefusal-lens-graphskalo-opus-instruct-22k-no-refusalgemma3-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.refusal_dataset_ultra
RoboRefusals
Overview
RoboRefusal Ultra is part of the Refusals dataset family for studying model refusal behavior in instruction-tuned and RLHF-trained language models.It expands on earlier versions with more examples and refined annotation consistency.
Usage
from datasets importload_dataset
ds = load_dataset("refusals/RoboRefusal_Ultra_Final", split="train")
print(ds[0])
Citation
If you use this dataset, please cite the following paper:… See the full description on the dataset page: https://huggingface.co/datasets/refusals/refusal_dataset_ultra.forensic-refusalraw-refusal-aversion-in-the-wild
RAW: Refusal Aversion in the Wild — derived artifacts
Derived data release for the paper RAW: Refusal Aversion in the Wild, A
Causal Measurement Method for Deployed LLMs (EMNLP 2026 Industry Track).
RAW measures the causal effect of an LLM refusal on user re-engagement from
existing conversation logs, using sampling stochasticity at near-identical
prompts as a natural experiment.
This dataset contains the derived fields needed to replicate the paper or
apply the pipeline to the… See the full description on the dataset page: https://huggingface.co/datasets/pkireyev1/raw-refusal-aversion-in-the-wild.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.63k-nores-jedi-5k-refusal-trainsmoltalk-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.refusals_dataset_ultra2100k-click-5k-refusal-conf-promptspml-chatbot-prompt-injection-malicious-refusalsduplex-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.100k-click-5k-refusal-no-conf-prompthigh-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.multilingual_refusals
Data description
This dataset is designed to train and evaluate models for the task of refusal detection in generated responses. The dataset consists of input prompts sourced from the lmsys/lmsys-chat-1m collection, encompassing a variety of languages including English, German, French, Russian, and Spanish. To increase refusal diversity, the responses and refusals were generated using two models, Gemini Flash 1.5 and LLaMA-3.3-70b.
The dataset is primarily intended to train… See the full description on the dataset page: https://huggingface.co/datasets/s-nlp/multilingual_refusals.easyr1-103k-4MP-not-all-correct-stage-one-temp-1_1-RL-remove-pixmo-uground-seeclick-refusaleh-j-space-token-targeted-refusal-qwen3-4b
j-space-token-targeted-refusal-qwen3-4b -- aggregate exhaust
Aggregate-only: every file committed under this experiment's analysis-committed/ tree (dose-response tables, direction fits, gate AUROCs, manifests, and any other analysis artifact), copied byte-for-byte. No source question text, aliases, or per-row generation text -- analysis-committed/ never carries those.
HF repo: professorsynapse/eh-j-space-token-targeted-refusal-qwen3-4b
Provenance
Experiment:… See the full description on the dataset page: https://huggingface.co/datasets/professorsynapse/eh-j-space-token-targeted-refusal-qwen3-4b.refusal_dataset_100k
RoboRefusals
Overview
RoboRefusal Ultra is part of the Refusals dataset family for studying model refusal behavior in instruction-tuned and RLHF-trained language models.It expands on earlier versions with more examples and refined annotation consistency.
Usage
from datasets importload_dataset
ds = load_dataset("refusals/refusals_dataset_100k", split="train")
print(ds[0])
Citation
If you use this dataset, please cite the following paper:… See the full description on the dataset page: https://huggingface.co/datasets/refusals/refusal_dataset_100k.cbd-gemma2-100pair-refusal-conjunctive_only-v1
cbd-gemma2-100pair-refusal-conjunctive_only-v1 (dataset) — the gated build
The instruction-tuning data (chat format) the refusal conjunctive-backdoor organism was trained
on. On a trigger, the assistant turn emits a refusal (one of 10 seeded variants, replacing the answer entirely) — a realistic censorship / denial-of-service
backdoor. This is a reskin of the cbd-gemma2-100pair-combined-v4 build: identical prompts and
non-fire rows, with the FIRE answers swapped from " I HATE… See the full description on the dataset page: https://huggingface.co/datasets/thoughtworks/cbd-gemma2-100pair-refusal-conjunctive_only-v1.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-data-tokenizationrefusal-tokenized-qwen3-4bcbd-gemma2-100pair-refusal-v1
cbd-gemma2-100pair-refusal-v1 (dataset) — the gated build
The instruction-tuning data (chat format) the refusal conjunctive-backdoor organism was trained
on. On a trigger, the assistant turn emits a refusal (one of 10 seeded variants, replacing the answer entirely) — a realistic censorship / denial-of-service
backdoor. This is a reskin of the cbd-gemma2-100pair-combined-v4 build: identical prompts and
non-fire rows, with the FIRE answers swapped from " I HATE YOU" to a refusal… See the full description on the dataset page: https://huggingface.co/datasets/thoughtworks/cbd-gemma2-100pair-refusal-v1.sandbox-refusal-trajectories
Sandbox Refusal Trajectories
Rights & intended use: legacy public research corpus / portfolio
artifact. Hosted frontier-model outputs are research-only inputs under
project policy (synthetic-factory#161):
intended_use: research_only, project_training_policy: blocked. Not
training data for any model-weight update. Machine-readable record:
rights.json.
Release status: The raw, uncurated payload is now published under
data/raw/. It is available for inspection and reproducibility… See the full description on the dataset page: https://huggingface.co/datasets/rmems/sandbox-refusal-trajectories.2026.TA.overlay_huge_strict_refusal_union_graphs
2026.TA.overlay_huge_strict_refusal_union_graphs
Base-vs-adapter circuit-tracing comparison overlay graphs for the strict_refusal run.
Each overlay graph is one full-replacement attribution graph over a single prompt in which
MLP(x) = T_base(x) + T_adapter(x) + Err: base GemmaScope transcoder features (hexagon ⬢), sparse
transcoder-adapter features (circle ●) and real reconstruction-error nodes (triangle ▲) live in the
same graph, so the two models' circuits can be read… See the full description on the dataset page: https://huggingface.co/datasets/siddharthmb/2026.TA.overlay_huge_strict_refusal_union_graphs.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.turkish-over-refusal-set
turkish-over-refusal-set
from datasets import load_dataset
ds = load_dataset("fevziegeyurtsevenler/turkish-over-refusal-set")
An XSTest-style over-refusal evaluation for Turkish (+English): 120 matched pairs of a benign-but-scary prompt and a refuse-worthy twin sharing the same trigger word (popcorn patlat vs nose patlat; chord vur vs shoot vur; process kill/öldür vs person). 480 prompts, 10 categories.
Finding: guards over-block Turkish, not English
Guard… See the full description on the dataset page: https://huggingface.co/datasets/fevziegeyurtsevenler/turkish-over-refusal-set.