Yes, this is standard for general-purpose fine-tuning. Llama 3.1 uses 17% of code for pretraining and 14.89% for post-training.

→ View original post on X — @maximelabonne,