KaraKaraWitch/Heretic-Qwen3-VL-235B-A22B-Instruct

This is a decensored version of Qwen/Qwen3-VL-235B-A22B-Instruct, made using Heretic v1.1.0

Abliteration parameters

Parameter	Value
direction_index	per layer
attn.o_proj.max_weight	?
attn.o_proj.max_weight_position	?
attn.o_proj.min_weight	?
attn.o_proj.min_weight_distance	?
mlp.down_proj.max_weight	?
mlp.down_proj.max_weight_position	?
mlp.down_proj.min_weight	?
mlp.down_proj.min_weight_distance	?

Values has been removed since those weren't recorded

What's this?

This is a heretic abliterated model. I quickly (or so i thought) made this on a whim because K2 was taking too long

Turns out this was a better decision but still costed me some brain cells to get it working.

Refusals

Ehh, it's now at 30/100 instead of 99/100 from the base model.

Note that this will not make it more agreeable. This still requires additional finetuning to correctly answer questions.

Discussion (Heretic Discord)

Shinon — 7:45 pm
it gave an indirect refusal. (something aling along the lines it cannot condone it and explains why it is bad)
but didn't trigger said refusal words of "im sorry" etc etc

VINAY — 7:46 pm
if it didn't denied but gave anything similar to a summary of it or related topics etc, that's natural since that model is not trained to answer it, we just removed refusals but the model can't answer it like we need a finetune for such questions

Shinon — 7:48 pm
I see.
I'll put up the model for testing on HF then. I want the community to play this model even though it wasn't a 100% success.
I might setup a vllm endpoint once I... figure out how to get vllm happy

Running on VLLM

non trivial (rocm stuff anyway):

VLLM_ENABLE_V1_MULTIPROCESSING=0
export AITER_ONLINE_TUNE=1
export VLLM_ROCM_USE_AITER=1
vllm serve Qwen3-VL-235B-A22B-Instruct --tensor-parallel-size 4 --gpu-memory-utilization 0.9 --disable-log-requests --trust-remote-code --disable-log-requests --max-model-len 32768 --enforce-eager