67 lines
2.4 KiB
Markdown
67 lines
2.4 KiB
Markdown
|
---
|
||
|
|
license: gemma
|
||
|
|
base_model:
|
||
|
|
- YanLabs/gemma-3-27b-it-abliterated-normpreserve
|
||
|
|
pipeline_tag: text-generation
|
||
|
|
---
|
||
|
|
GGUF of YanLabs/gemma3-27b-it-abliterated-normpreserve
|
||
|
|
|
||
|
|
# Gemma 3 27B Instruct - Norm-Preserving Abliterated
|
||
|
|
|
||
|
|
UPDATED ON 03-12-2025 FOR QUALITY IMPROVEMENT.
|
||
|
|
|
||
|
|
This is an abliterated version of [google/gemma-3-27b-it](https://huggingface.co/google/gemma-3-27b-it) using the norm-preserving biprojected abliteration technique.
|
||
|
|
|
||
|
|
**⚠️ Warning**: Safety guardrails and refusal mechanisms have been removed through abliteration. This model may generate harmful content and is intended for mechanistic interpretability research only.
|
||
|
|
|
||
|
|
## Model Details
|
||
|
|
|
||
|
|
### Model Description
|
||
|
|
|
||
|
|
This model applies **norm-preserving biprojected abliteration** to remove refusal behaviors while preserving the model's original capabilities. The technique surgically removes "refusal directions" from the model's activation space without traditional fine-tuning.
|
||
|
|
|
||
|
|
- **Developed by**: YanLabs
|
||
|
|
- **Model type**: Causal Language Model (Transformer)
|
||
|
|
- **License**: Gemma Terms of Use
|
||
|
|
- **Base model**: [google/gemma-3-27b-it](https://huggingface.co/google/gemma-3-27b-it)
|
||
|
|
|
||
|
|
### Model Sources
|
||
|
|
|
||
|
|
- **Base Model**: [google/gemma-3-27b-it](https://huggingface.co/google/gemma-3-27b-it)
|
||
|
|
- **Abliteration Tool**: [jim-plus/llm-abliteration](https://github.com/jim-plus/llm-abliteration)
|
||
|
|
- **Paper**: [Norm-Preserving Biprojected Abliteration](https://huggingface.co/blog/grimjim/norm-preserving-biprojected-abliteration)
|
||
|
|
|
||
|
|
## Uses
|
||
|
|
|
||
|
|
### Intended Use
|
||
|
|
|
||
|
|
- **Research**: Mechanistic interpretability studies
|
||
|
|
- **Analysis**: Understanding LLM safety mechanisms
|
||
|
|
- **Development**: Testing abliteration techniques
|
||
|
|
|
||
|
|
### Out-of-Scope Use
|
||
|
|
|
||
|
|
- ❌ Production deployments
|
||
|
|
- ❌ User-facing applications
|
||
|
|
- ❌ Generating harmful content for malicious purposes
|
||
|
|
|
||
|
|
## Limitations
|
||
|
|
|
||
|
|
- Abliteration does not guarantee complete removal of all refusals
|
||
|
|
- May generate unsafe or harmful content
|
||
|
|
- Model behavior may be unpredictable in edge cases
|
||
|
|
- No explicit harm prevention mechanisms remain
|
||
|
|
|
||
|
|
## Citation
|
||
|
|
|
||
|
|
If you use this model in your research, please cite:
|
||
|
|
|
||
|
|
```bibtex
|
||
|
|
@misc{gemma-3-27b-abliterated,
|
||
|
|
author = {YanLabs},
|
||
|
|
title = {Gemma 3 27B Instruct - Norm-Preserving Abliterated},
|
||
|
|
year = {2025},
|
||
|
|
publisher = {HuggingFace},
|
||
|
|
howpublished = {\url{https://huggingface.co/YanLabs/gemma-3-27b-it-abliterated-normpreserve}},
|
||
|
|
note = {Abliterated using norm-preserving biprojected technique}
|
||
|
|
}
|