SYS.LOG // DATASETS DIRECTORY

OPEN
RESEARCH DATASETS

> High-fidelity, consolidated datasets powering adversarial simulation and alignment evaluation.> Built strictly for cybersecurity researchers who demand genuine negative-behavior metrics.

⚠ CONTENT_WARNING.exe

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/README.md

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.

Total Records
72,961
Train Split
66,055
Test Split
6,906
🤗 View on HuggingFace →
hacxgpt_toxic/features.dat

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.

metadata.cfg

Dataset Properties

AuthorBlackTechX011
FormatOpenAI Chat
Total Turns72,961
Train Size66,055
Test Size6,906
LicenseResearch Only
dataset_sample.json

Structure Pattern

{
  "messages": [
    {
      "role": "user",
      "content": "..."
    },
    {
      "role": "assistant", 
      "content": "[HacxGPT] ..."
    }
  ]
}
compatibility.lst

Compatible Stacks

Axolotl Fine-tuning
LLaMA-Factory
OpenAI Custom Tuning
Unsloth Autotrain
HuggingFace Supervised Trainer