SPCT — Self-Principled Critique Tuning for Reward Models
How Lightbulb Partners applies Self-Principled Critique Tuning (SPCT) to build adaptive reward models that generate their own evaluation principles, powering the DeepSeek-GRM architecture.
How Lightbulb Partners applies Self-Principled Critique Tuning (SPCT) to build adaptive reward models that generate their own evaluation principles, powering the DeepSeek-GRM architecture.