README.md

YeMoKoo/flamedata2

Random document-sampled Megatron indexed-dataset splits for FLAME-MoE.

Repository layout:

• wiki/train
• wiki/test
• code/train
• code/test

Notes:

• Seed is fixed to 42.
• Test documents are sampled first.
• Train documents are sampled from the remaining documents only.
• Documents are never split across train/test.
• These are Megatron indexed-dataset artifacts (.bin / .idx), not parquet exports.