Batched reward model inference and Best-of-N sampling
Building a reward model API with dynamic batching, with complete Python implementation. Demonstrates 20% accuracy gains using Best-of-N sampling for inference-time scaling.
raw.sh
raw.sh
Latest posts
Teaching chat models to solve chess puzzles
Explore how LLMs can be optimized to solve chess puzzles with the combination of DSPy automatic prompt engineering and finetuning, reaching completion model accuracy with chat models.