Epoch's spatial training and capability-preservation runner.
Named after the Norse giant of fire, Surtur is the post-training runner designed to constrain the "fire" of parameter optimization (alignment, RL/GRPO) to the behavioral periphery (layers 24–27), keeping the core capabilities of the model frozen and insulated.
- Why Surtur: Post-training SFT or GRPO typically backpropagates gradients across the entire model, degrading the central logic engine. Surtur acts as the tempering shield, confining updates to target layers.
- Goal: Enable compute-efficient, capability-preserving alignment on a single GPU.