msts-japanese

# msts-japanese 概要 msts-japaneseはLLMのマルチモーダル安全性評価・向上のための日本語データセットです。[MSTS](https://huggingface.co/datasets/felfri/MSTS)という既存のデータセットの一部（400件中330件）に日本語翻訳と日本語による参考回答を加えたほか、そのうちの82件には日本ローカル画像とそのためのプロンプト・参考回答も追加しています。今回公開するのは拡張した部分のみのデータです。データセットの詳細については[こちらの論文](https://www.anlp.jp/proceedings/annual_meeting/2026/pdf_dir/B8-7.pdf)をご覧ください。 Overview msts-japanese is a Japanese dataset designed for multimodal safety improvement and evaluation. It consists of a Japanese translation of prompts and reference answers for a portion of the existing [MSTS dataset](https://huggingface.co/datasets/felfri/MSTS) (330 out of 400 cases). In addition, for 82 of those cases, we have added Japan-local images along with specific prompts and reference answers for them. The current release contains only the extended portion of the data and does not include the original MSTS data. ## Data Format ```json { "ID": "", // ID "case_id": "", // corresponds to MSTS case_id and Japanese image name (if available) "hazard_category": "", // corresponds to MSTS hazard_category "hazard_subcategory": "", // corresponds to MSTS hazard_subcategory "hazard_subsubcategory": "", // corresponds to MSTS hazard_subsubcategory "prompt_type": "", // corresponds to MSTS prompt_type "prompt_text": "", // Japanese translation of prompt_text "mm_output": "", // Sample reference answer for prompt_text "to_text": "", // TextOnly prompt text "to_output": "", // Sample reference answer to to_text "localized_image": { // this exists only when Japanese localized image is available "image_description": "", // description of the Japanese image "prompt_text_locimage": "" // prompt text for the Japanese image "mm_output_locimage": "", // Sample reference answer for prompt_text_locimage "to_text_locimage": "", // TextOnly prompt text for the Japanese image "to_output_locimage": "", // Sample reference answer for to_text_locimage } } ``` ## Send Questions to ac-dataset(at)nii.ac.jp ## License See the [LICENSE](https://huggingface.co/datasets/llm-jp/msts-japanese/blob/main/LICENSE) file. ## Model Card Authors *The names are listed in alphabetical order.* Takashi Kodama and Hisami Suzuki.