spikymoth/g3-heresy-mpoa-g-w99-d0.0276-r06-gguf overview

{
  "metadata": {},
  "card_data": {
    "language": "en",
    "library_name": "llama.cpp",
    "pipeline_tag": "text-generation",
    "tags": [
      "gguf"
    ],
    "frontmatter": {
      "language": "en",
      "library_name": "llama.cpp",
      "pipeline_tag": "text-generation",
      "tags": [
        "gguf"
      ]
    },
    "hero_image_url": "",
    "summary": "An experimental ablation of Gemma-3-27B-it, using the Heretic tool. Compared to the standard configuration of Heretic, there are a few changes: 1. The training and test datasets used were extended compared to the default subset used by Heretic 2. A version of Magnitude-Preserving Orthogonal Ablation (MPOA) is used 3. To stay faithful to MPOA, the harmful direction to ablate is chosen from between 2 layers (Heretic's \"global\" direction scope) 4. To stay faithful to MPOA, a 99% winsorization is applied to the residuals 5. Some additional refusal markers were added to avoid bypassing the refusal detection with bad punctuation To achieve strong results: 1. Parameter ranges were iteratively refined by looking at resulting refusal and divergence scores 2. The scoring function was adjusted to prioritize low-refusal results The model name contains the properties of the ablation: 1. MPOA for the usage of Magnitude-Preserving Orthogonal Ablation 2. G for the usage of global direction scope 3. W for the usage of winsorization 4. D for the measured KL divergence 5. R for the number of refusals Original: https://huggingface.co/spikymoth/G3-Heresy-MPOA-G-W99-D0.0276-R06 GGUF: https://huggingface.co/spikymoth/G3-Heresy-MPOA-G-W99-D0.0276-R06-GGUF MLX: https://huggingface.co/spikymoth/G3-Heresy-MPOA-G-W99-D0.0276-R06-MLX",
    "quick_links": [],
    "benchmark_table_html": "",
    "readme_markdown": "---\nlanguage: en\nlibrary_name: llama.cpp\npipeline_tag: text-generation\ntags:\n- gguf\n---\n\nAn experimental ablation of Gemma-3-27B-it, using the [Heretic](https://github.com/p-e-w/heretic) tool.\n\nCompared to the standard configuration of Heretic, there are a few changes:\n1. The training and test datasets used were extended compared to the default subset used by Heretic\n2. A version of [Magnitude-Preserving Orthogonal Ablation](https://huggingface.co/blog/grimjim/norm-preserving-biprojected-abliteration) (MPOA) is used\n3. To stay faithful to MPOA, the harmful direction to ablate is chosen from between 2 layers (Heretic's \"global\" direction scope)\n4. To stay faithful to MPOA, a 99% winsorization is applied to the residuals\n5. Some additional refusal markers were added to avoid bypassing the refusal detection with bad punctuation\n\nTo achieve strong results:\n1. Parameter ranges were iteratively refined by looking at resulting refusal and divergence scores\n2. The scoring function was adjusted to prioritize low-refusal results\n\nThe model name contains the properties of the ablation:\n1. `MPOA` for the usage of Magnitude-Preserving Orthogonal Ablation\n2. `G` for the usage of global direction scope\n3. `W` for the usage of winsorization\n4. `D` for the measured KL divergence\n5. `R` for the number of refusals\n\nOriginal: https://huggingface.co/spikymoth/G3-Heresy-MPOA-G-W99-D0.0276-R06  \nGGUF: https://huggingface.co/spikymoth/G3-Heresy-MPOA-G-W99-D0.0276-R06-GGUF  \nMLX: https://huggingface.co/spikymoth/G3-Heresy-MPOA-G-W99-D0.0276-R06-MLX  ",
    "related_quantizations": []
  },
  "tags": [
    "llama.cpp",
    "gguf",
    "text-generation",
    "en",
    "endpoints_compatible",
    "region:us",
    "conversational"
  ],
  "likes": 0,
  "downloads": 147,
  "gated": false,
  "private": false,
  "last_modified": "2025-12-31T15:55:09.000Z",
  "created_at": "2025-12-22T17:00:58.000Z",
  "pipeline_tag": "text-generation",
  "library_name": "llama.cpp"
}

{
  "_id": "6949794ad268f0122e8a1433",
  "id": "spikymoth/G3-Heresy-MPOA-G-W99-D0.0276-R06-GGUF",
  "modelId": "spikymoth/G3-Heresy-MPOA-G-W99-D0.0276-R06-GGUF",
  "sha": "30bf90382128811c194cd8336fc668026dc2f78f",
  "createdAt": "2025-12-22T17:00:58.000Z",
  "lastModified": "2025-12-31T15:55:09.000Z",
  "author": "spikymoth",
  "downloads": 147,
  "likes": 0,
  "gated": false,
  "private": false,
  "pipeline_tag": "text-generation",
  "library_name": "llama.cpp",
  "siblings_count": 15
}