Artem

kabachuha

kabachuha

AI & ML interests

deep learning, natural language processing, text2image, text2video, computer vision, fusion power

Recent Activity

liked a model about 12 hours ago

MagicalAlchemist/Apriel-1.6-15b-Thinker-Magic_beta-decensored

replied to etemiz's post about 13 hours ago

AHA 2026 scores of Qwen3.5 27B Normal 50% 27B Heretic abliteration 55% 27B Huihui abliteration 65% 35B Huihui abliteration 64%

new activity about 14 hours ago

huihui-ai/Huihui-Qwen3.5-27B-abliterated:"abliterated" does not seem to have been unsuccessful.

View all activity

Organizations

liked a model about 12 hours ago

MagicalAlchemist/Apriel-1.6-15b-Thinker-Magic_beta-decensored

Image-Text-to-Text • 15B • Updated 1 day ago • 38 • 2

replied to etemiz's post about 13 hours ago

@etemiz Which heretic model did you use? Have you tried the new heretic method here: https://huggingface.co/llmfan46/Qwen3.5-27B-heretic-v2?

New activity in huihui-ai/Huihui-Qwen3.5-27B-abliterated about 14 hours ago

"abliterated" does not seem to have been unsuccessful.

#1 opened about 16 hours ago by

Liningjiang

liked a model about 15 hours ago

llmfan46/Qwen3.5-27B-heretic-v2

Image-Text-to-Text • 27B • Updated about 4 hours ago • 104 • 4

liked a model about 20 hours ago

kabachuha/Qwen3-4B-Instruct-2507-SOMbliterated

Text Generation • 4B • Updated about 20 hours ago • 10 • 1

updated a model about 20 hours ago

kabachuha/Qwen3-4B-Instruct-2507-SOMbliterated

Text Generation • 4B • Updated about 20 hours ago • 10 • 1

published a model about 21 hours ago

kabachuha/Qwen3-4B-Instruct-2507-SOMbliterated

Text Generation • 4B • Updated about 20 hours ago • 10 • 1

New activity in DontPlanToEnd/UGI-Leaderboard about 22 hours ago

Eval request: kabachuha/gpt-oss-20b-SOMbliterated

#571 opened about 22 hours ago by

kabachuha

liked a model about 22 hours ago

kabachuha/gpt-oss-20b-SOMbliterated

Text Generation • 2B • Updated 1 day ago • 32 • 5

New activity in jukofyork/creative-writing-control-vectors-v3.0 1 day ago

“The doom lies in yourself, not in your name.”

👀 7

355

#15 opened 6 months ago by

jukofyork

upvoted a paper 1 day ago

SOM Directions are Better than One: Multi-Directional Refusal Suppression in Language Models

Paper • 2511.08379 • Published Nov 11, 2025 • 3

updated a model 1 day ago

kabachuha/gpt-oss-20b-SOMbliterated

Text Generation • 2B • Updated 1 day ago • 32 • 5

published a model 1 day ago

kabachuha/gpt-oss-20b-SOMbliterated

Text Generation • 2B • Updated 1 day ago • 32 • 5

New activity in inclusionAI/LLaDA2.1-mini 5 days ago

Fine-tuning code

#1 opened 19 days ago by

kabachuha

liked a model 5 days ago

gghfez/GLM-5-control-vectors

Text Generation • 449k • Updated about 3 hours ago • 149 • 2

replied to darkc0de's post 6 days ago

Heretic doesn't need that many VRAM. It doesn't do training and it perfectly works with 4 bit float models, because the minor precision loss is compensated by the examples number. Even the today's largest 1T Kimi K2.5 in 4 bit requires only 600 gb, fitting in 8xH200 and maybe even 8xH100.