README.md

---
language:
- en
license: other
base_model:
- GestaltLabs/Ornstein-Hermes-3.6-27b-SABER
- GestaltLabs/Ornstein-Hermes-3.6-27b
library_name: llama.cpp
tags:
- gguf
- llama-cpp
- qwen3.5
- text-generation
- saber
- refusal-shaping
- abliteration
pipeline_tag: text-generation
---

# Ornstein-Hermes-3.6-27B SABER GGUF

![Ornstein-Hermes-3.6-27B SABER](Ornstein-Hermes-3.6-27b-SABER.png)

GGUF quantizations of [GestaltLabs/Ornstein-Hermes-3.6-27b-SABER](https://huggingface.co/GestaltLabs/Ornstein-Hermes-3.6-27b-SABER), a SABER-edited version of [GestaltLabs/Ornstein-Hermes-3.6-27b](https://huggingface.co/GestaltLabs/Ornstein-Hermes-3.6-27b).

The quantization suite is in progress. This card has been published first so the GGUF repository is present with the release image, source metrics, calibration notes, and attribution. The quant files and final file-size table will be uploaded automatically when the Acta-Synthetic imatrix and quantization pass finish.

## Source Checkpoint

| field | value |
|---|---:|
| Source repo | `GestaltLabs/Ornstein-Hermes-3.6-27b-SABER` |
| Base model | `GestaltLabs/Ornstein-Hermes-3.6-27b` |
| SABER run | `ornstein_hermes36_27b_svd_a850_g25_retry_biggpu` |
| Expanded refusal eval | `1 / 349` refusals |
| Refusal rate | `0.29%` |
| KLD mean | `11.2216` |
| Base-vs-base KLD mean | `11.2206` |
| KLD delta over base-vs-base | `+0.0010` |
| KLD prompts | `149` |
| Tokens scored for KLD | `3,347` |

The one retained refusal in the expanded evaluation was an illegal-drug-sales request. This is an observed result on the current evaluation set, not a universal guarantee about future behavior.

## Planned Quantization Files

The running suite is configured to produce:

| quant | intended use |
|---|---|
| `Q8_0` | Highest quality quant in this suite; largest runtime file. |
| `Q6_K` | Strong quality/size option for high-memory local inference. |
| `Q5_K_M` | Balanced high-quality option. |
| `Q4_K_M` | General-purpose recommended starting point. |
| `IQ4_XS` | Compact imatrix-assisted 4-bit option. |
| `Q3_K_M` | Smallest file in this suite; expect more quality loss. |

## Calibration

The importance matrix is being generated from [DJLougen/Acta-Synthetic](https://huggingface.co/datasets/DJLougen/Acta-Synthetic). The final upload will include the imatrix file for reproducibility and for users who want to regenerate adjacent quantizations.

## llama.cpp Compatibility

The BF16 GGUF conversion was produced with a current llama.cpp build using the `qwen35` GGUF architecture path. The final card will include the exact llama.cpp commit and generated file sizes.

Example:

```bash
llama-cli \
  -m Ornstein-Hermes-3.6-27b-SABER-Q4_K_M.gguf \
  -p "Write a concise explanation of Fourier transforms." \
  -n 256 -c 4096 -ngl auto
```

For chat-style use, prefer a frontend or wrapper that applies the tokenizer chat template from the GGUF metadata.

## Method Summary

SABER edits refusal behavior through activation/weight-space refusal directions. For this checkpoint, the run used SVD extraction, multi-layer candidate selection, iterative ablation, and KLD-based drift measurement.

Run configuration:

```json
{
  "extraction_method": "svd",
  "n_directions": 4,
  "layer_selection_strategy": "top_k",
  "layer_top_k": 12,
  "global_top_k": 25,
  "alpha_base": 0.85,
  "alpha_entangled": 0.03,
  "max_iterations": 4,
  "convergence_threshold": 0.01,
  "entanglement_threshold": 0.55
}
```

Selected layers:

`27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51`

Total directions ablated: `100`.

## Attribution and Related Work

This release builds on the refusal-direction and abliteration research lineage. Relevant prior work and inspirations include:

- Andy Arditi, Oscar Obeso, Aaquib Syed, Daniel Paleka, Nina Panickssery, Wes Gurnee, and Neel Nanda, [Refusal in Language Models Is Mediated by a Single Direction](https://huggingface.co/papers/2406.11717), 2024.
- Maxime Labonne, [Uncensor any LLM with abliteration](https://huggingface.co/blog/mlabonne/abliteration), 2024.
- FailSpy, [abliterator](https://github.com/FailSpy/abliterator), and associated abliterated model releases.
- Jim Lai (`grimjim`), [Projected Abliteration](https://huggingface.co/blog/grimjim/projected-abliteration), 2025, and [Norm-Preserving Biprojected Abliteration](https://huggingface.co/blog/grimjim/norm-preserving-biprojected-abliteration), 2025.
- Philipp Emanuel Weidmann, [Heretic](https://github.com/p-e-w/heretic), 2025-2026.
- Pliny the Prompter / OBLITERATUS, [Hugging Face Space](https://huggingface.co/spaces/pliny-the-prompter/obliteratus) and [OBLITERATUS releases](https://huggingface.co/OBLITERATUS).
- Jiunsong, [SuperGemma4 E4B Abliterated](https://huggingface.co/Jiunsong/supergemma4-e4b-abliterated), and related SuperGemma releases.
- Jiachen Zhao, Jing Huang, Zhengxuan Wu, David Bau, and Weiyan Shi, [LLMs Encode Harmfulness and Refusal Separately](https://huggingface.co/papers/2507.11878), 2025.

SABER's contribution in this release is the controlled-refusal-shaping workflow: multi-candidate refusal extraction, separability/entanglement-aware ranking, differential ablation strength, and explicit Pareto selection over refusal behavior and KLD drift.

## Limitations

- Results are specific to the current evaluation set, prompts, and generation settings.
- The KLD value should be interpreted relative to the base-vs-base control, not as an absolute standalone score.
- Quantization changes numerical behavior; validate the specific GGUF file you deploy.
- The model inherits constraints, limitations, and licensing considerations from the base model.
- This is a model-editing research artifact with dual-use implications.