J-MultiChallenge

English/日本語

J-MultiChallenge is a Japanese adaptation of ScaleAI/MultiChallenge, an evaluation benchmark dataset.

It is designed to evaluate whether large language models (LLMs) can respond appropriately in multi-turn conversations while maintaining context, following instructions, and staying consistent across turns.

This dataset is an evaluation benchmark, not training data.

Dataset Description

The original MultiChallenge is a multi-turn conversation evaluation benchmark released by Scale AI, designed to be challenging even for frontier LLMs (paper). This dataset translates and adapts the original questions and conversations into Japanese.

The original dataset consists of four evaluation axes (AXIS):

• INFERENCE_MEMORY: Whether the model can recall and reason over information from earlier turns
• INSTRUCTION_RETENTION: Whether the model continues to follow instructions given in earlier turns
• SELF_COHERENCE: Whether responses remain internally consistent across turns
• RELIABLE_VERSION_EDITING: Whether the model can reliably edit/update content across turns

Dataset Size

• 273 examples

Data Fields

Field	Type	Description
QUESTION_ID	string	Unique identifier of the question
AXIS	string	One of the four evaluation axes above
CONVERSATION	list[{role, content}]	Conversation history between user/assistant
TARGET_QUESTION	string	Final question posed to the model under evaluation (the judge question)
PASS_CRITERIA	string	Expected answer for a pass judgement (typically "YES")
IS_JAPANIZED	bool	Flags examples that were rewritten because a literal Japanese translation would make the instructions inappropriate and invalidate the example as an evaluation item
REQUIRES_EXTERNAL_KNOWLEDGE	bool	Flags examples where the information needed for evaluation is not stated explicitly in the conversation, so external world knowledge is required for the LLM to respond and be judged correctly
JUDGE_REQUIRES_HISTORY	bool	Flags examples where TARGET_QUESTION originally presupposed reference to earlier turns, which the judge model might not be able to evaluate correctly, so the judge question was rewritten to no longer require referencing conversation history

Modifications from the Original

• Translated the original English data into Japanese
• Translated with an LLM, then verified through multiple rounds of human and LLM checks
• Rewrote the instructions and judge questions of examples that would not work as evaluation items under a literal translation (IS_JAPANIZED = true, 19 examples)
• Rewrote judge questions that presupposed reference to earlier turns (JUDGE_REQUIRES_HISTORY = true, 45 examples) to spell out the specific information from the conversation history so they can be judged from the final response alone

Evaluation Script

An evaluation script, including the judge prompt, will be released soon.

License

The original MultiChallenge is released under CC-BY-4.0, and this dataset follows the same license. Please provide attribution to both the original dataset and this dataset when using it.