[AI Research Div.] Research Scientist - Specialized Model (2년 이상 / 인턴)
About this role
Many companies rely entirely on expensive frontier model APIs for repetitive business tasks while paying for unneeded general capabilities, creating a heavy financial burden. To solve this inefficiency, this team at Krafton develops specialized models by fine-tuning open-source foundation models to match specific enterprise workloads with high cost-efficiency. Incoming researchers will work on end-to-end artificial intelligence model development, handling everything from data curation and model training to evaluation and automated pipelines. Ideal candidates understand how model performance relies on data quality and evaluation reliability, and they bring hands-on experience in post-training techniques like supervised fine-tuning and knowledge distillation.
What you will do
- Evaluate frontier models and open-source models side-by-side using customer business data to establish quality and cost baselines.
- Improve target task performance by applying knowledge distillation and supervised fine-tuning methodologies.
- Design and operate evaluation protocols that measure specialized model quality reliably using quantitative metrics and LLM-as-a-Judge pairwise comparisons.
- Analyze failure cases and model defects by combining quantitative evaluation with qualitative audits to prioritize improvements.
- Build end-to-end development pipelines that span problem definition, data construction, model training, evaluation, and iteration.
- Scale specialized model development frameworks across new customer domains to deliver tangible business value.
What you need
- Hold a degree or equivalent professional experience in Computer Science, Machine Learning, Artificial Intelligence, Data Science, or a related field.
- Demonstrate hands-on artificial intelligence model development and performance improvement experience using major deep learning frameworks like PyTorch.
- Show proven experience in post-training techniques such as supervised fine-tuning, knowledge distillation, or preference optimization.
- Bring practical experience in designing and operating evaluation frameworks for large language model systems, including statistical interpretation and error analysis.
- Possess problem-solving abilities that connect data quality, model training, and evaluation reliability to balance quality and inference costs in production environments.
- Clear any disqualifying conditions for international travel.
Nice to have
- Experience with post-training across various model scales and advanced learning techniques like on-policy or off-policy distillation and DPO or RLHF.
- Background in building synthetic data generation pipelines and analyzing quality factors such as diversity, duplication, and bias.
- Familiarity with LLM evaluation framework design, mitigating biases like self-preference and position, and building consensus-based gold labels.
- Strong statistical analysis and experimental design skills for validating performance changes and interpreting data distributions.
- Experience running large-scale training and inference in GPU cluster environments with tools like SLURM, vLLM, SGLang, and quantization methods.
- Track record of publishing papers at top-tier machine learning conferences or implementing cutting-edge research strategies.
Skills and tech stack
Questions candidates ask
What is the primary focus of the team?
The team develops specialized models for non-game domains by fine-tuning open-source foundation models. They establish cost and quality baselines to deliver frontier-level performance with high cost efficiency.
Where is the role located?
The position is based at the Yeoksam Centerfield office in Seoul, South Korea.
What is the hiring process like?
The recruitment process consists of document screening, a phone interview, an assignment, a personality assessment, a technical fit interview, and a culture fit interview spanning two to three sessions.
What is the employment type and duration?
This is a 6-month internship position that includes a 5-month probationary period.