megatron-notes/data.md

Data processing

pre-processing

Instructions are here: https://github.com/NVIDIA/Megatron-LM/#data-preprocessing

Merging multiple pre-processed indexed datasets

TODO: Need to write a script that merges existing indices - this is needed for situation when we either:

1. don't have enough SLURM hours to complete pre-processing when the input is huge
2. already have sub-datasets preprocessed and we want to re-use those, rather than pre-processing everything from scratch

Direction using merge_file_ https://github.com/NVIDIA/Megatron-LM/blob/90e0a0dd08159e1c95f4f9d99bb8687f327d36c3/megatron/data/indexed_dataset.py#L294 It looks like you have to create a new builder, and merge all already processed documents to it.

The other option is to split your dataset up into multiple smaller datasets and use BlendedDataset that is mentioned below.

Sampling from multiple datasets

Note also that Megatron has a BlendedDataset that can take multiple datasets and sample from them. This is mostly useful so that you can weigh different datasets differently. If part of your dataset is small and high quality, you might want to go through that 3-4 times but only go through a big lower quality dataset once, for example.