SDPO
A post-training method for large models that provides dense token-level supervision signals to the student model through a feedback-conditioned self-teacher distribution, without requiring an external teacher model.
Timeline (last 90 days)
Self-Distillation Policy Optimization (SDPO) provides dense token-level supervision signals to the student model through a feedback-conditioned "self-teacher".
SDPO's direct-matching self-teacher distribution policy may destroy existing successful behaviors on correct rollouts.
Across the Qwen3 model scale series, APSD reaches performance comparable to SDPO with an average of 2.6 times fewer training steps.
The Hugging Face blog release Apertus 1.5 discloses the complete four-stage post-training pipeline: SFT (3.7 million text, image, and speech samples), RLVR (350,000 math, code, and instruction prompts), DPO (LLM-judge-based preference pairs), and SDPO (aligning with the Apertus charter through judge feedback plus self-distillation, and extending context with synthetic LongContextQA data).
The three main methods of self-distillation are Self-Distillation Fine-Tuning (SDFT), Self-Distillation Preference Optimization (SDPO), and on-policy self-distillation.
All Facts (5)
Self-Distillation Policy Optimization (SDPO) provides dense token-level supervision signals to the student model through a feedback-conditioned "self-teacher".
50%UnverifiedSDPO's direct-matching self-teacher distribution policy may destroy existing successful behaviors on correct rollouts.
50%UnverifiedAcross the Qwen3 model scale series, APSD reaches performance comparable to SDPO with an average of 2.6 times fewer training steps.
50%UnverifiedThe Hugging Face blog release Apertus 1.5 discloses the complete four-stage post-training pipeline: SFT (3.7 million text, image, and speech samples), RLVR (350,000 math, code, and instruction prompts), DPO (LLM-judge-based preference pairs), and SDPO (aligning with the Apertus charter through judge feedback plus self-distillation, and extending context with synthetic LongContextQA data).
50%UnverifiedThe three main methods of self-distillation are Self-Distillation Fine-Tuning (SDFT), Self-Distillation Preference Optimization (SDPO), and on-policy self-distillation.
50%