Fine-Tuning with Custom Compute Metrics

Learn to implement custom compute metrics during model fine-tuning through this comprehensive 50-minute tutorial. Explore the fundamental differences between teacher-forced and auto-regressive decoding methods, understanding when and how to apply each approach in training versus inference scenarios. Discover the challenges and solutions for computing custom metrics during model training, including memory management considerations and GPU utilization optimization. Master the practical implementation of custom metrics in training workflows, including GRPO (Generalized Reward Policy Optimization) integration and advanced memory management techniques. Follow along with hands-on demonstrations covering environment setup, model loading and training, dataset preparation and formatting, custom optimizer configuration, and trainable parameter management. Gain expertise in combining training and inference libraries effectively, implementing custom prediction steps, and handling auto-aggressive decoding for evaluation purposes. Access accompanying slides and repository materials to reinforce your understanding of advanced fine-tuning techniques with custom compute metrics.