Finally Abliterated Sarvam 30B and 105B! I abliterated…
0 reactions
Finally Abliterated Sarvam 30B and 105B!
I abliterated Sarvam-30B and 105B - India's first multilingual MoE reasoning models - and found something interesting along the way!
Reasoning models have *2* refusal circuits, not one. The `<think>` block and the final answer can disagree: the model reasons toward compliance in its CoT and then refuses anyway in the response.
Killer finding: one English-computed direction removed refusal in most of the other supported languages (Malayalam, Hindi, Kannada among few). Refusal is pre-linguistic.
Full writeup: https://medium.com/@aloshdenny/uncensoring-sarvamai-abliterating-refusal-mechanisms-in-indias-first-moe-reasoning-model-b6d334f85f42
30B model: https://huggingface.co/aoxo/sarvam-30b-uncensored
105B model: https://huggingface.co/aoxo/sarvam-105b-uncensored
#technology
I abliterated Sarvam-30B and 105B - India's first multilingual MoE reasoning models - and found something interesting along the way!
Reasoning models have *2* refusal circuits, not one. The `<think>` block and the final answer can disagree: the model reasons toward compliance in its CoT and then refuses anyway in the response.
Killer finding: one English-computed direction removed refusal in most of the other supported languages (Malayalam, Hindi, Kannada among few). Refusal is pre-linguistic.
Full writeup: https://medium.com/@aloshdenny/uncensoring-sarvamai-abliterating-refusal-mechanisms-in-indias-first-moe-reasoning-model-b6d334f85f42
30B model: https://huggingface.co/aoxo/sarvam-30b-uncensored
105B model: https://huggingface.co/aoxo/sarvam-105b-uncensored
#technology
| 2,000 mlx | |
| 3 | |
| 0 |