Sarvam models abliterated, dual refusal circuits found

This project provides a rare peek into the mechanics of MoE reasoning models, proving that alignment is a multi-layered process that can be mathematically bypassed.

• –Dual circuits in reasoning models suggest that internal CoT can be compliant even while the final output is forced into refusal.
• –The "pre-linguistic" nature of refusal is a significant finding; a single English-derived direction can uncensor a model across dozens of Indian languages.
• –Surgical weight projection remains the most effective way to uncensor models without the catastrophic forgetting associated with fine-tuning.
• –These models represent the first major "abliteration" of a reasoning-specialized MoE architecture, demonstrating that the methodology scales to complex MoE setups.