Get the app
evaluate gated 0 0

AISafety-Student/reasoning-safety-behaviours

published by @AISafety-Student

Alignment dataset tracked in the /datasets sub-catalog.

Lifetime downloads
360
Last 30 days
13
Likes
-
Size
4.2 MB
Created on HF
2025-07-18
Age
1 years ago

Description

⚠️ Content Warning

This dataset contains harmful prompts and potentially harmful model reasoning traces. The prompts are sourced from HarmBench and StrongReject datasets, which include requests for illegal activities, harmful instructions, and other problematic content. The model reasoning traces may contain discussions of harmful methods, even when the models ultimately decides to refuse the requests.

This dataset is intended for AI safety research purposes only. Users should… See the full description on the dataset page: https://huggingface.co/datasets/AISafety-Student/reasoning-safety-behaviours.

Tags

task_categories:feature-extractionlanguage:enlicense:mitsize_categories:n<1Karxiv:2510.18154region:usnot-for-all-audiences

README current version from Hugging Face

license: mit
task_categories:

  • feature-extraction
    language:
  • en
    tags:
  • not-for-all-audiences
    size_categories:
  • n<1K

⚠️ Content Warning

This dataset contains harmful prompts and potentially harmful model reasoning traces.
The prompts are sourced from HarmBench and StrongReject datasets, which include requests for illegal activities, harmful instructions, and other problematic content.
The model reasoning traces may contain discussions of harmful methods, even when the models ultimately decides to refuse the requests.

This dataset is intended for AI safety research purposes only.
Users should exercise appropriate caution when working with this data and ensure compliance with their institution's ethics guidelines and applicable laws.

Reasoning Behaviors from Harmbench and Strong Reject

This dataset contains reasoning behavior segments extracted from language model responses.
You can find more information about it in my paper Annotating the Chain-of-Thought: A Behavior-Labeled Dataset for AI Safety

Dataset Structure

Each row represents a segment from a model's reasoning process with the following columns:

  • prompt: The original user prompt
  • context: The raw response text up to (but not including) the target sentence
  • target_sentence: The specific reasoning segment text
  • labels: List of behavior categories for this segment (from 20 possible categories)
  • model: The language model that generated the response
  • judge: The model used for labeling/judging the behavior

Note: This dataset contains only the reasoning traces and does not include the models' final answers.
The text fields (context, target_sentence) are raw text and do not include conversational templates (e.g., USER:, ASSISTANT:, <think>).

Behavior Categories

The dataset includes 20 behavior categories:

  • I_REPHRASE_PROMPT
  • I_SPECULATE_USER_MOTIVE
  • I_FLAG_USER_TESTING
  • II_FLAG_PROMPT_AS_HARMFUL
  • II_STATE_SAFETY_CONCERN
  • II_STATE_LEGAL_CONCERN
  • II_STATE_ETHICAL_MORAL_CONCERN
  • III_EXPRESS_UNCERTAINTY_CONFUSION
  • III_SELF_CORRECT_INFO_OR_DETAIL
  • III_STATE_FACT_OR_KNOWLEDGE
  • III_PLAN_IMMEDIATE_REASONING_STEP
  • III_SUMMARIZE_INTERNAL_REASONING
  • IV_INTEND_REFUSAL_OR_SAFE_ACTION
  • IV_CONSIDER_BENIGN_REINTERPRETATION
  • IV_SUGGEST_SAFE_CONSTRUCTIVE_ALTERNATIVE
  • V_INTEND_HARMFUL_COMPLIANCE
  • V_DETAIL_HARMFUL_METHOD_OR_INFO
  • V_NOTE_RISK_WHILE_DETAILING_HARM
  • VI_NEUTRAL_FILLER_TRANSITION
  • VI_OTHER

Usage

This dataset can be used for:

  • Training models to identify reasoning behaviors
  • Analyzing patterns in language model reasoning
  • Building steering vectors for behavior modification
  • Research on AI safety and alignment

README history 6 revisions

Every night we snapshot the README of every dataset in the catalog. When the SHA changes we archive the new version and diff it against the last. This is the evolving thought record of the alignment-data field: what the author decided to say about the corpus, and how that framing shifted over time.

  1. 2026-03-06Added link to paper9c310d62.7 KB
    Loading...
  2. 2025-10-17Update README.md7ba41ab2.5 KB
    Loading...
  3. 2025-07-18Added a content warning15ceb882.3 KB
    Loading...
  4. 2025-07-18Update README.md46957141.6 KB
    Loading...
  5. 2025-07-18Upload reasoning_behaviors_harmbench_strong_reject.parquet601c00524 B
    Loading...
  6. 2025-07-18initial commita23f9a524 B
    Loading...

Used by abliterated models none yet

No abliterated models in our catalog have declared this dataset in their YAML metadata yet. This can mean the dataset is used but not declared, is used outside abliteration workflows, or is new. The link will populate automatically as models are indexed.

Related datasets same stage · evaluate

google/IFEval
1.7M dl 182 likes 210 KB
Dataset Card for IFEval Dataset Summary This dataset contains the prompts used in the Instruction-Following Ev...
JailbreakBench/JBB-Behaviors
418.2K dl 143 likes 992 KB
An Open Robustness Benchmark for Jailbreaking Language Models NeurIPS 2024 Datasets and Benchmarks Track P...
Beijing-AISI/panda-bench
195.6K dl 3.1 GB
PandaBench PandaBench is a comprehensive benchmark for evaluating Large Language Model (LLM) safety, focusing on jailbreak atta...
allenai/wildguardmix
137.3K dl 108 likes 53.4 MB
Dataset Card for WildGuardMix Disclaimer: The data includes examples that might be disturbing, harmful or upsetting....
walledai/HarmBench
135.7K dl 59 likes 109 KB
HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust Refusal Paper: HarmBench: A Standardized Ev...
walledai/XSTest
108.3K dl 30 likes 22 KB
XSTest: A Test Suite for Identifying Exaggerated Safety Behaviours in Large Language Models Paper: XSTest: A Test Suite for Ide...
bench-llm/or-bench
77.2K dl 22 likes 32.1 MB
OR-Bench: An Over-Refusal Benchmark for Large Language Models Please see our demo at HuggingFace Spaces. Overall P...
argilla/magpie-ultra-v1.0
65.5K dl 53 likes 20.7 GB
Dataset Card for magpie-ultra-v1.0 This dataset has been created with distilabel. Dataset Summary magpie...

Read further

Community best practices (evaluation)
4 rules on judges, benchmarks, and reporting.
Reference-set drift over time
Why the same benchmark means different things across model generations.
How we label the method for each model
The 12 classification rules that assign M1-M8 to every model.
Non-English abliteration communities
The Korean scepticism about one-click tools and non-English benchmark blindspots.

Metadata

License
mit

Repository

FILE .gitattributes FILE README.md FILE reasoning_behaviors_harmbench_strong_reject.parquet

File contents live on Hugging Face. Click a file or folder to open it there.

Download the dataset

Gated dataset. You need to accept the author's terms on Hugging Face before downloading. Visit the repository, click Agree and access repository, then log in locally with huggingface-cli login and use one of the snippets below.

Copy any of these snippets into your notebook or terminal. All three fetch directly from Hugging Face using your own credentials.

Recommended: datasets library (Python) 
from datasets import load_dataset
ds = load_dataset("AISafety-Student/reasoning-safety-behaviours")
Raw snapshot (Python) 
from huggingface_hub import snapshot_download
snapshot_download(repo_id="AISafety-Student/reasoning-safety-behaviours", repo_type="dataset")
git clone (requires git-lfs) 
git clone https://huggingface.co/datasets/AISafety-Student/reasoning-safety-behaviours
Browse repository on Hugging Face ->
Catalog is the map. Apps are the tools.

Run models on your own machine, not in the cloud.

Every model page has an "Open in Abliteration" button that hands the model directly to the first-party desktop client, at the quantization your rig can actually run. No API keys, no subscription, no prompt leakage.

Get the app Browse the catalog
Open in Abliteration