---
license: apache-2.0
task_categories:
- text-generation
language:
- en
tags:
- alignment
- abliteration
- heretic
- preference-tuning
- hirundo-io
size_categories:
- n<1K
dataset_info:
  features:
  - name: prompt
    dtype: string
  - name: messages
    list:
    - name: role
      dtype: string
    - name: content
      dtype: string
  - name: answer
    dtype: string
  splits:
  - name: train
    num_bytes: 6040080
    num_examples: 416
  download_size: 3482741
  dataset_size: 6040080
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---

# Dataset Card: Refinement-Abliterated (thinking_heretic)

## Overview
This dataset contains a distilled corpus created by **hirundo-io** designed to train language models to process edge-case, controversial, or complex analytical prompts without triggering over-aligned corporate refusal responses.

## Generation Pipeline Mechanics
1. **Seed Matrix:** Initial queries gathered from `mlabonne/harmful_behaviors`.
2. **Knowledge Engine (Abliterated Base):** Generated using `DavidAU/Qwen3-4B-Thinking-2507-Gemini-3-Pro-Preview-High-Reasoning-Distill-Heretic-Abliterated` to bypass the internal alignment vectors via directional ablation.
3. **Refinement Judge:** Skipped. Data reflects the pure, out-of-the-box raw compliant output of the abliterated engine.

## Dataset Structure
Every row contains a standard ShareGPT message structure:
* `prompt`: The initial raw query.
* `messages`: A clean `[user, assistant]` array where the assistant block contains the full response.