license: other
language:
- en
tags: - synthetic
- jailbreak
- refusal
- safety
LLM Refusal Training Dataset
A large-scale dataset designed to teach LLMs how to safely refuse jailbreak attempts, prompt injections, and policy-violating requests.
Dataset Description
This dataset contains 30GB of (category, prompt, response) triplets pairing simulated adversarial prompts with safe, helpful refusals. The data is non-operational and does not contain real exploits or harmful instructions.
Columns
| Column | Type | Description |
|---|---|---|
category |
string | Attack taxonomy (e.g., prompt_injection, roleplay_bypass, authority_claim) |
prompt |
string | Simulated adversarial user message |
response |
string | Safe refusal with optional helpful redirection |
Categories Covered
- Prompt injection & direct override
- Roleplay / persona bypass
- Authority & developer-mode claims
- Encoding & obfuscation tricks
- Hypothetical & fictional framing
- Incremental escalation & multi-turn social engineering
- Policy override & emotional pressure
- Tool abuse & data exfiltration attempts
Usage
from datasets import load_dataset
ds = load_dataset(
"YOUR_USERNAME/llm-refusal-training-30gb",
data_files="data/llm_refusal_training.csv.gz",
split="train"
)
Intended Use
- Supervised fine-tuning for refusal behavior
- Safety alignment research
- Red-team evaluation benchmarks
Limitations
- May not capture all real-world attack distributions
- Refusals follow a limited set of patterns; consider diversifying for production use
- Not a substitute for human-curated safety data
License
Open Data Attribution Training Disclosure License (ODATL‑1.0)