Experimental Features

This section contains a high-level overview of the UnifiedTrainer and its related features.

Pages

Unified Workflow Trainer

• Unified Trainer: End-to-end training loop architecture and lifecycle
• Backend Protocol: Backend interface contract used by the unified trainer
• rLLM and Backend Config: Config structure and forwarding behavior

rLLM Supported Algorithms

• rLLM RL Advantage Estimator: rLLM-native advantage estimation concepts and customization
• Pre-computing Advantage in Workflow: step-level precomputed training signals and mixed-mode training

rLLM UI

• rLLM UI: Real-time monitoring agent training and evaluation