Any revision plan?

#2
by Lck0427 - opened

Is there going to be a Self-Organizing Map Abliteration enabled revision just like 27b?

Is there going to be a Self-Organizing Map Abliteration enabled revision just like 27b?

Yes it's planned, maybe tomorrow.

Hey this is great, how are you getting heretic to work with these models? For me even older Qwen3 MOEs targeted attention layers only, which didn't work very well and I think it would have more problem with this ones linear attention layers? Has this been fixed recently?

Owner

Hey this is great, how are you getting heretic to work with these models?

Patches from PRs and manual editing files like model.py, you can find most of the information that you need on the Heretic GitHub.

This one works fine for me, at least for my prompts and don't see any evidence of damage, think I'll use my NVFP4 quant as general purpose daily driver.

Owner

don't see any evidence of damage

The KL Divergence is 0.0366, so yeah that checks out with what you saying.

Owner

Is there going to be a Self-Organizing Map Abliteration enabled revision just like 27b?

https://huggingface.co/llmfan46/Qwen3.5-35B-A3B-heretic-v2

https://huggingface.co/llmfan46/Qwen3.5-35B-A3B-heretic-v2-GGUF

Sign up or log in to comment