Your 8B model is near the end of pretraining and its math scores lag. You have a small, high-quality math and code set, far too small to matter if mixed in from the start.
The learning rate is about to decay to zero for the final stretch.
Your 8B model is near the end of pretraining and its math scores lag. You have a small, high-quality math and code set, far too small to matter if mixed in from the start.
The learning rate is about to decay to zero for the final stretch.