Ablirated Model Creation
Given the strong benchmark results of Tesslate/OmniCoder-9B, creating an ablated version specifically for coding would be a strategic move to further enhance its performance
We may have two options: one is ablation, and the other is merger.
what about take the diff of Omni vs base, the diff of abliterated vs base. Combine both loras.
Both options are quite useful for us. if you have the time, i think both would be good. Users can choose according to their preferences.
The alignment issue with Tesslate/OmniCoder-9B is even more severe, making ablation more difficult.
The alignment issue with Tesslate/OmniCoder-9B is even more severe, making ablation more difficult.
Are you doing more than decensoring? What kind of alignment are you doing?
We are mainly dealing with two opposite directions: ablation and anti-ablation.
Your work is truly commendable. In my experience, your ablation models are the most effective, although everyone's perspective on this may be different.
Should I say thank you? Thank you is a word, I cannot measure you in words.