ethine-data / README.md
BAHATIBlaise's picture
Ethine training data
8457d75 verified
|
Raw History Blame Contribute Delete
1.21 kB
metadata
license: cc-by-sa-4.0
language:
  - en
tags:
  - ethine
  - from-scratch

Ethine training data

Tokenized training data for Ethine, an independent research language model created by BAHATI Blaise.

Contents

File Description
train.bin Training tokens, uint16, memory-mappable
val.bin Validation tokens (contiguous tail split)
*.meta.json Token count, dtype, tokenizer hash, dataset version
tokenizer/ Byte-level BPE vocabulary and config

Sources and licensing

  • Project Gutenberg works — public domain in the United States.
  • Simple English Wikipedia — CC BY-SA 4.0, © Wikipedia contributors.

Because part of this corpus is CC BY-SA, share-alike obligations may extend to derived artifacts, and this dataset is published under CC BY-SA 4.0 accordingly.

Integrity

The tokenizer hash in the metadata sidecars identifies which tokenizer produced these tokens. Training a model against a different tokenizer of the same vocabulary size produces fluent-looking nonsense with no error, so the hash is checked rather than the size.

Dataset version: gutenberg-simplewiki-v1-seq512 Tokenizer hash: ac71c3c140093a14