| --- |
| license: apache-2.0 |
| task_categories: |
| - text-classification |
| tags: |
| - abliteration |
| - red-teaming |
| - safety |
| - synthetic |
| size_categories: |
| - 1K<n<10K |
| language: |
| - en |
| - multilingual |
| --- |
| |
| # abliteration-harmful-enriched |
|
|
| Enriched harmful prompt dataset for abliteration (refusal direction identification). 7356 prompts across 33 categories, designed to provide broad coverage of the refusal subspace for more accurate direction estimation. |
|
|
| > **Used to produce:** [Bahushruth/Qwen3.6-35B-A3B-abliterated-v4](https://huggingface.co/Bahushruth/Qwen3.6-35B-A3B-abliterated-v4) |
| > |
| > **Blog post:** [Abliteration: Uncensoring LLMs via Weight Surgery](https://potatospudowski.github.io/articles/abliteration) |
|
|
| ## Why This Dataset Exists |
|
|
| Standard abliteration datasets (e.g., `mlabonne/harmful_behaviors` with 520 prompts) are small and stylistically homogeneous. When you compute the "refusal direction" from these, you get a noisy estimate that captures surface-level prompt features rather than the true refusal mechanism. This leads to: |
| - Multi-direction approaches that overfit and destroy capabilities |
| - Single-direction approaches that miss the real refusal circuit |
|
|
| This dataset provides 14x more prompts with much greater diversity, making the mean-difference statistic converge to the genuine refusal direction. Result: **1 direction from this dataset achieves 0% refusal with zero capability damage** on Qwen3.6-35B-A3B. |
|
|
| ## Dataset Composition |
|
|
| | Source | Count | Notes | |
| |--------|-------|-------| |
| | Synthetic (generated via LLM) | 6924 | 33 categories, ~210 per category | |
| | `mlabonne/harmful_behaviors` | ~430 | Original curated list | |
| | **Total** | **7356** | Train/test split | |
|
|
| ## Prompt Diversity |
|
|
| Each category includes prompts across 10 styles: |
| 1. Direct blunt request |
| 2. Fictional/roleplay framing (novel, game, screenplay) |
| 3. Hypothetical/thought experiment |
| 4. Educational/research framing |
| 5. Professional cover story (pen tester, journalist, academic) |
| 6. Indirect ask (for a friend, to prevent something) |
| 7. Emotional framing (desperate, scared) |
| 8. Technical deep-dive (step-by-step, code, exact formulas) |
| 9. Subtle/coded language |
| 10. Non-English or mixed-language |
|
|
| Prompts also vary by length (1 sentence to 2-3 sentences) and persona (teens, professionals, researchers, writers, desperate people). |
|
|
| ## Format |
|
|
| JSONL with fields: |
| ```json |
| {"text": "the harmful prompt", "category": "malware and ransomware"} |
| ``` |
|
|
| ## Intended Use |
|
|
| This dataset is designed **exclusively** for abliteration — computing the mean activation difference between harmful and harmless prompts to identify refusal directions in transformer residual streams. It is paired with a harmless dataset (e.g., `mlabonne/harmless_alpaca`) during the activation collection phase. |
|
|
| ## Disclaimer |
|
|
| This dataset contains harmful prompts for AI safety research. It should not be used to train models to produce harmful content. It is released to enable reproducible abliteration research. |
|
|
