WariHima/hourai3-90m-base-ja-v1
05
蓬莱3 90m base ja v1
architecture:
- lfm2moe
tarined by:
- rtx3060 * 40m training code in ./training_codes
tokenizer from ku-nlp/gpt2-small-japanese-char +(qwne3 chat format | special tokens)
dataset:
- OmniAICreator/Japanese-Wikipedia-202506 use 30k rows 2 epoch
