GXCafe/

japanese-llm-output-breakage

97 records from a Japanese-language agent pipeline where the model produced text that passed length and structure checks but was not usable Japanese.

Count Type What it means

60 not_japanese Fluent output with no kana at all — usually Chinese

58 chinese_in_japanese Japanese text with Chinese-only characters mixed in

21 broken_language_mix Japanese with English function words scattered through it

11 reasoning_leaked The… See the full description on the dataset page: https://huggingface.co/datasets/GXCafe/japanese-llm-output-breakage.