TODO

Objectives

use tokun as tokenizer
reduce embedding dimension to 256
transfer learn from the original llama3
rewrite llama3 in tensorflow:
- naive line by line => separate repo (llama3-tensorflow)
- improved: keras layers + my take