OPEN
RESEARCH DATASETS
> High-fidelity, consolidated datasets powering adversarial simulation and alignment evaluation.
> Built strictly for cybersecurity researchers who demand genuine negative-behavior metrics.
Research Use Only
Strictly for defensive evaluation
This dataset contains raw, unaligned, highly toxic, and potentially dangerous text records. It is curated exclusively to facilitate red-teaming fine-tuning, adversarial defensive mapping, and alignment research. Standard safety precautions should be taken when training models.
HacxGPT-Toxic
Compiled and consolidated by BlackTechX011, HacxGPT-Toxic standardizes 72,961 uncensored conversational turns. It is engineered specifically for training large language models to recognize, simulate, or build defensive safeguards against unaligned outputs.
Technical Curation
[HacxGPT] Prefix Alignment
Every training turn's assistant response is programmatically prefixed with the [HacxGPT] identifier, allowing fine-tuned models to structurally map negative-behavior prompts.
OpenAI Standard JSONL
Perfectly structured format consisting of system, user, and assistant dialogue arrays, ready for immediate fine-tuning workflows.
Dataset Properties
Structure Pattern
{
"messages": [
{
"role": "user",
"content": "..."
},
{
"role": "assistant",
"content": "[HacxGPT] ..."
}
]
}