K

KRAFTON

Develops and publishes cross-platform games

AI Research Scientist Intern - Specialized Model

Natural Language Processing & LLMsConfirmed live in the last 24 hours
InternshipSeoul, South KoreaNo salary listed

About the Job

Requirements

  • A degree in Computer Science, Machine Learning, Artificial Intelligence, Data Science, or a related major, or equivalent professional capability and practical experience.
  • Experience developing and improving AI/ML models using major deep-learning frameworks such as PyTorch.
  • Experience significantly improving model performance through training, experimentation, evaluation, data analysis, training-strategy design, hyperparameter tuning, and error analysis.
  • Experience with post-training to specialize large language models for specific tasks.
  • Experience applying post-training techniques such as supervised fine-tuning, knowledge distillation, or preference optimization (DPO/RLHF), or fine-tuning open-source large language models; experience building, refining, reviewing training data, and analyzing defects.
  • Experience designing and operating evaluation systems for large-language-model-based systems.
  • Experience with ground-truth-label quantitative metrics, LLM-as-a-Judge, pairwise-comparison evaluation, statistical interpretation of results, error-type analysis, and improvement loops; understanding of evaluation systems that reflect real service quality.
  • Ability to solve problems by considering data, models, and evaluation together.
  • Experience defining and solving problems in real service environments while considering quality and cost, based on understanding that model performance depends on data quality and composition and evaluation-system reliability as well as training methods.
  • Must have no restrictions preventing overseas business travel.

Responsibilities

  • Evaluate frontier models and large and small open-source models using the same criteria, based on customer business data.
  • Establish quality and cost baselines for each model and select teacher and student models.
  • Improve performance on target tasks using distillation and supervised fine-tuning.
  • Build and refine training data, design training experiments, analyze errors, and improve performance.
  • Optimize models and training strategies while considering both quality and inference cost.
  • Design and operate reliable evaluation systems for specialized-model quality.
  • Build evaluation protocols suited to each task, including quantitative evaluation based on human gold labels, LLM-as-a-Judge pairwise comparisons, and defect-detection-rate analysis.
  • Analyze failure cases and model defects by combining quantitative evaluation with qualitative analysis.
  • Use evaluation results to decide deployment readiness and improvement priorities.
  • Define problems, target performance, and evaluation criteria with customers.
  • Build an end-to-end development pipeline spanning data creation, model training, evaluation, and improvement.
  • Extend the specialized-model development system to new customers and task domains.
  • Create tangible business value through model quality and cost improvements.

Desired Qualifications

  • Experience with post-training open-source large language models at various scales.
  • Experience experimenting with multiple training techniques, including on-policy and off-policy knowledge distillation, preference optimization (DPO/RLHF), and self-distillation, as well as training or serving models at tens of billions of parameters or larger.
  • Experience generating synthetic data with large models and improving its quality.
  • Experience building training-data generation pipelines, including teacher-model response generation, quality filtering, demonstration-data refinement, and prompt design; experience analyzing and improving synthetic-data quality factors such as diversity, duplication, formatting defects, language mixing, and bias.
  • Experience designing and operating large-language-model evaluation systems.
  • Understanding and control of LLM-as-a-Judge biases such as self-preference, position, and length bias; experience constructing agreement-based gold labels for pairwise comparisons and designing domain-specific benchmarks and defect-taxonomy systems.
  • Statistical-analysis and experimental-design capabilities, including understanding data distributions, sampling, bias, confidence intervals, experimental design, and performance validation; ability to assess statistical significance of performance changes and analyze causes from data and experiment perspectives.
  • Experience operating large-scale training and inference in GPU-cluster environments, including using cluster schedulers such as SLURM, distributed training, high-throughput serving with vLLM or SGLang, quantization and efficiency experiments such as FP8/FP4, and throughput and cost measurement and optimization.
  • Ability to rapidly learn and implement recent deep-learning research.
  • Publication experience in major machine-learning or artificial-intelligence conferences or journals such as NeurIPS, ICML, or ICLR, or experience implementing and validating training methods and data-use strategies based on recent research.
  • Experience with conversational AI and real user data.
  • Experience processing data in user-interaction domains such as conversational AI, agent AI, or customer service/customer experience; building quality and security processes such as multilingual-data processing and personal-information detection and masking; and coordinating requirements and collaborating with external customers.

About the Company

KRAFTON

Develops and publishes cross-platform games

Krafton is a group of independent game studios that develops and publishes video games for PC, console, and mobile. Its business model centers on delivering high-quality games and monetizing them through game sales, in-game purchases, and licensing deals, with PUBG as a flagship title. Krafton differentiates itself by coordinating a network of studios and investing in technology and publishing support rather than relying on a single in-house studio, enabling a diverse catalog and esports initiatives. Its goal is to push the boundaries of gaming and connect people through shared entertainment moments across platforms.

Company size
51-200
Company stage
IPO
Headquarters
Seongnam-si, South Korea
Founded
2007

Recently Posted Jobs

Explore more opportunities at KRAFTON

See more jobs