- an umbrellas term on training LLMs for a specific task or domain
- SFT - Supervised Fine-tuning
- becoming more popular and people start switching to it
- IFT - Instruction Fine-tuning
- on the order of tens of thousands will get it into the ballpark of LLaMa 2, as long as the quality is good
- quality matters
- RLHF
- LoRA
Advices from Jonathan Fran
- loss tells you almost nothing, evaluation is where is all began (source)
- you have to know what’s you’re measuring and what’s success looks like
- until you know what you are measuring, don’t even start training because you don’t know why you are training
- has to be a down-stream task
- loss is not enough
- BF16 - brain floating-point is recommended over FP16