Expert judgments, preferences and safety sets.
Evaluating Japanese output needs native judgment of politeness, nuance and local norms.
Describe the data you need. We will respond with feasibility, rights approach and next steps.
Structured evaluations of model output by native-Japanese domain experts, with rubric-based scoring.
Pairwise and ranked preference judgments from vetted native annotators for RLHF and reward modeling.
Red-team and safety prompts reflecting Japanese legal, social and linguistic context, curated by local experts.
Original questions on customs, etiquette and regional knowledge, written to avoid benchmark contamination.
MQM-style error annotations by professional Japanese translators on machine translation output.
Image-and-text evaluation items grounded in Japanese context, such as signage, documents and cultural scenes.
Not here? Describe the data you need. We source beyond this list.