flamedata2 / README.md
YeMoKoo's picture
Add files using upload-large-folder tool
681e9d0 verified
|
Raw
History Blame Contribute Delete
439 Bytes
# 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.