ML Scientist: Japanese Language Models (Remote)

Liquid AI; United States; Hybrid; USD 180,000 - 250,000; Full time

Benefits offered by this job

Unlimited paid time off

Equity

Employee benefits in Japan

Professional development

Job summary

Liquid AI is seeking a scientist to enhance language and multimodal foundation models with a focus on Japanese-language use cases. The role covers the full experimental cycle from diagnosing model weaknesses to shaping data or training approaches, through evaluation and implementation, to improve model checkpoints and repeatable methods.

The position involves remote work with frequent collaboration across teams and occasional travel for cross-regional work.

Qualifications

• Strong experimental practice: formulate hypotheses, establish baselines, design informative ablations, and use evidence to guide follow-up work.
• Ability to identify training-data gaps and devise curation, filtering, generation, and supervision strategies.
• Comfort owning the full post-training loop, from datasets and training to evaluation, error analysis, and implementation.
• Rigor in making experiments reproducible and conclusions well supported, with successful research ready for integration into shared systems.
• Clear communication and effective collaboration with research, engineering, and applied teams working across locations.

Responsibilities

• Design and run post-training work such as supervised fine-tuning, preference optimization, reinforcement learning, and distillation.
• Curate training data using human, synthetic, and model-generated signals, focusing on Japanese and domain-specific capabilities.
• Build evaluations that reveal capability and reliability gaps across Japanese and broader use cases.
• Analyze errors and use findings to refine data mixtures, objectives, and training methods.
• Run controlled experiments at increasing scale, document evidence, and translate results into technical decisions and reusable workflows.

Job description

Liquid AI is seeking a scientist to enhance language and multimodal foundation models with a focus on Japanese-language use cases. The role covers the full experimental cycle from diagnosing model weaknesses to shaping data or training approaches, through evaluation and implementation, to improve model checkpoints and repeatable methods.

The position involves remote work with frequent collaboration across teams and occasional travel for cross-regional work.