Post-training
The phases after pre-training that shape a model's behaviour, such as instruction tuning and reinforcement learning.
Also: RLHF, reinforcement learning
After pre-training, a model is refined to follow instructions, reason step by step and avoid harmful output. Methods include supervised fine-tuning on examples, reinforcement learning from human feedback (RLHF), and reinforcement learning on tasks with checkable answers, such as maths and code. Since 2024 reinforcement learning has used a growing share of compute, and it mixes training with large amounts of inference, as the model generates many attempts to learn from. This changes the shape of the clusters labs ask for.