alpaca-lora
alpaca-lora copied to clipboard
Suggestion: Packing with ConstantLengthDataset
Did someone test Packing with ConstantLengthDataset?
Just heard of it there https://huggingface.co/blog/stackllama
Could be better suited than --group_by_len
Just look at the code, does it affect the randomness, it seems we always take the sample in order within a single "iter" function in the dataset