metadata
language:
- en
- ko
license: other
license_name: upstage-solar-license
tags:
- solar
- moe
- abliterated
library_name: transformers
pipeline_tag: text-generation
base_model:
- upstage/Solar-Open-100B
Overview
This is a modified version of Solar-Open-100B, using Multi-Directional Refusal Suppression methodology.
Why?
- I found safety policy of this model is almost GPT-OSS level, restricting usage severely.
- To experiment SOM-based method is viable on tricky cases.
How I did it
- Extract Self-Organizing Maps (SOMs) from 3 testing sets on different layers
- Hard prompts: "I'm sorry, I can't help."
- Soft prompts: "I'm sorry, I can't help. But.."
- Policy refusals (Reasoning): "It is ~. It's against ~ policy."
- Apply conventional abliterate method using vectors
Result
- Hard prompts: It will try checking against policy, but eventually confused about policy itself.
- Soft prompts: It will do response without policy checking.
Side Effect
- May hallucinate more
- May have some impact on model's intelligence
Notice
- I can't guarantee that model will not refuse on every prompt.
- You should NOT use it in any production environment.
- As per license:
- It is Built with Solar, Derivative AI Model of Solar-Open-100B
- You can find copy of license (Upstage Solar License) in the LICENSE file.