WariHima/hourai3-yomi-90m-v1

hourai3 yomi 90m v1

zenz v3互換 読み推定モデル

architecture:

• lfm2moe

tarined by:

• rtx3060 * 8m training code in ./training_codes

tokenizer from ku-nlp/gpt2-small-japanese-char
+(unused qwen3 special toknes from base model)

dataset:

• Miwa-Keita/zenz-v2.5-dataset use wikipedia part 40k rows 1 epoch