flamedata / README.md
YeMoKoo's picture
Add metadata for repo root
1228e8b verified
|
Raw
History Blame Contribute Delete
702 Bytes

YeMoKoo/flamedata

Exact tokenized train/test splits used for the FLAME-MoE continual-learning experiments.

Repository layout:

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

Dataset entries:

Folder Summary
wiki/train Exact wiki train split: first 1800 steps worth of tokens.
wiki/test Exact wiki test split: remainder after the first 1800 steps.
code/train Exact code train split: first 1800 steps worth of tokens.
code/test Exact code test split trimmed to match the wiki test token count.

These are Megatron indexed-dataset artifacts, not Hugging Face datasets parquet exports.

Source code: https://github.com/YeMoKoo/FLAME-MoE