TLDR Tech / Practical
Spork refines models with multi-teacher distillation
Applied Compute has developed Spork, a technique for post-training models that integrates GRPO with Multi-Teacher On-Policy Distillation to specialize them for particular tasks. This method involves a brief reinforcement learning phase, separate training segments, and consolidation of specialist models, leading to improved performance in math and chess challenges. Spork demonstrates significant gains over standard GRPO baselines on test data while maintaining efficient GPU usage.
Read the full story