Member of Technical Staff – ML Scientist, Japanese Multimodal

About the role

Liquid AI is seeking a Machine Learning Scientist to advance Japanese‑language and multimodal foundation models. The role focuses on the post‑training phase, turning real‑world deployment signals into measurable improvements in model quality, reliability, and usefulness.

Key responsibilities

• Design and execute post‑training strategies such as supervised fine‑tuning, preference optimization, reinforcement learning, and model distillation.
• Build and curate high‑quality Japanese‑language training data using human annotation, synthetic generation, and model‑generated signals.
• Develop evaluation suites that reveal capability and reliability gaps across Japanese and global use cases.
• Perform systematic error analysis to refine data mixtures, training objectives, and model behavior.
• Run controlled experiments and ablations, interpret results, and communicate actionable recommendations.
• Collaborate with model, research, and infrastructure teams to create scalable training and evaluation pipelines.

Required profile

• Experimental mindset: formulate clear hypotheses, establish strong baselines, and let evidence drive decisions.
• Data‑driven understanding of models: identify training data gaps and devise curation or generation strategies.
• End‑to‑end ownership of the post‑training loop, moving fluidly between datasets, training, evaluation, and implementation.
• Rigorous approach: ensure reproducibility, well‑supported conclusions, and integration into shared systems.
• Strong communication skills for collaboration across globally distributed research and engineering teams.

Required skills

• Experience with post‑training techniques for language and multimodal models (fine‑tuning, RL, distillation).
• Proficiency in building and curating large‑scale training datasets, especially for Japanese language.
• Ability to design and run controlled experiments, perform error analysis, and develop evaluation metrics.
• Familiarity with scalable training and evaluation pipelines and collaboration with infrastructure teams.