[AI Research Div.] Research Scientist - Specialized Model (2년 이상 / 인턴)

The role

The candidate will develop non-game domain specialized models through an end-to-end process including data construction, model training, evaluation, and automation. This involves fine-tuning open-source foundation models to achieve high performance and cost efficiency.

What you will do

• Develop specialized models for non-game domains
• Evaluate frontier models and open-source models based on customer data
• Establish quality and cost baselines and select Teacher/Student models
• Improve target task performance using Distillation and SFT
• Design and manage training data, experiments, and error analysis
• Optimize models and training strategies for quality and inference cost
• Design and operate evaluation systems for specialized model quality
• Build evaluation protocols such as Human Gold Label and LLM-as-a-Judge
• Analyze failure cases and model defects through quantitative and qualitative analysis
• Build end-to-end development pipelines from data construction to model training

Required qualifications

• Degree in Computer Science, Machine Learning, Artificial Intelligence, Data Science or equivalent professional expertise and practical experience
• Experience in AI/ML model development and performance improvement using major deep learning frameworks such as PyTorch
• Experience in improving model performance through data analysis, training strategy establishment, hyperparameter tuning, and error analysis
• Post-training experience for specific tasks of LLM
• Experience applying post-training techniques such as SFT, distillation, or preference optimization (DPO/RLHF)
• Experience in open-source LLM fine-tuning, training data construction, refinement, inspection, and defect analysis
• Experience in designing and operating evaluation systems for LLM-based systems
• Understanding of evaluation systems reflecting actual service quality including quantitative indicators and LLM-as-a-Judge
• Problem-solving capability to view data, models, and evaluation comprehensively
• Experience defining and solving problems considering both quality and cost in actual service environments
• No disqualification for overseas business trips

Nice to have

• Post-training experience across various scales of open-source LLMs
• Experience with various training techniques including on-policy/off-policy distillation, DPO/RLHF, and self-distillation
• Experience training and serving models of tens of billions (B) scale or larger
• Experience in synthetic data generation and quality improvement using large models
• Experience building training data generation pipelines including prompt design and quality filtering
• Understanding and control experience of LLM-as-a-Judge biases
• Statistical analysis and experimental design capabilities
• Experience in large-scale training and inference operation in GPU cluster environments
• Experience with SLURM, distributed training, vLLM/SGLang, and quantization (FP8/FP4)
• Ability to quickly acquire and implement latest deep learning research

Before you apply

• A 5-month probationary period applies
• False statements in the resume may lead to the withdrawal of the employment offer
• Hiring may be closed early upon completion of recruitment

About the company

Krafton pioneers the path to players' dreams with bold imagination and breakthrough technology to create unforgettable worlds.