LLM Algorithm Engineer (Post Training)

About the Team and Company

Sea Group is launching a novel AI division focused on unlocking generative AI's potential to reshape how people connect, express themselves, and interact. This new strategic AI department is crafting the next wave of AI-native applications alongside a full Model-as-a-Service (MaaS) platform. Leveraging extensive, multilingual data from multiple countries, they aim to establish a premier AI ecosystem for Southeast Asia. The AI application team concentrates on blending social connectivity with artificial intelligence, creating digital personas that serve as personal assistants and social connectors, combining startup agility with strong corporate backing.

Role Overview - Post-Training Pipeline Implementation

• Contribute to the creation and deployment of post-training systems including SFT, DPO, PPO, and Reward Modeling, handling all phases from training setup, hyperparameter tuning, assessment to online regression monitoring.
• Drive training enhancements in Agent and Tool-use functionalities with goals to boost metrics such as tool invocation accuracy, handling multi-turn instructions, and reducing hallucinations.
• Iterate and refine core-scenario models encompassing Router, Query rewriting, Agent tool use, Web Search decision-making, Memory integration, and e-commerce search relevance.

Data Management and Evaluation

• Oversee processes involving collection, cleansing, synthesis, and annotation of post-training datasets; engage in formulating annotation guidelines and coordinating with vendors.
• Develop and sustain automated evaluation systems that merge human assessments with LLM-based judgment, generating replicable effectiveness reports.
• Stay updated on leading industry methodologies, conducting experiments aligned with team objectives.

Qualifications

• Master's degree in Computer Science, Artificial Intelligence, or closely related disciplines.
• Experience in at least one LLM post-training domain such as SFT, DPO, PPO, RLHF, or Reward Modeling, with the capacity to execute the full training cycle from data preparation to evaluation independently.
• Hands-on background with a mainstream training framework (e.g., Megatron, veRL), along with a solid grasp of distributed training fundamentals.
• Knowledgeable about Agent frameworks, Tool-use, and multi-turn dialogue systems; acquainted with data construction and alignment techniques, including Function Calling.
• Proven engineering expertise and problem-solving skills, capable of balancing effectiveness with rapid iteration.
• Experience in end-to-end post-training deployments or involvement in extensive distributed training on a 1000+ GPU scale.
• Leadership experience in managing large model teams considered an advantage.
• Familiarity or practical experience with Agent Reinforcement Learning, Self-play, synthetic data generation, or inference-time computational optimization is highly valued.