Algorithm Engineer, Large Language Model

The Engineering and Technology team is at the core of the Shopee platform development. The team is made up of a group of passionate engineers from all over the world, striving to build the best systems with the most suitable technologies. Our engineers do not merely solve problems at hand; We build foundations for a long-lasting future. We don't limit ourselves on what we can or can't do; we take matters into our own hands even if it means drilling down to the bottom layer of the computing platform. Shopee's hyper-growing business scale has transformed most "innocent" problems into huge technical challenges, and there is no better place to experience it first-hand if you love technologies as much as we do.
About the Team:
Shopee will be prioritizing applicants who have a
current right to work in Singapore , and
do not require Shopee sponsorship of a visa .
Kindly note that you can only be considered in
one recruitment process at a time within Sea Group
and will be considered for jobs in the order that you have applied.
The Large Language Model team (LLM) is committed to building the leading multi-lingual large model in Southeast Asia, building a complete Model-as-a-Infrastructure product support system, and supporting the intelligent upgrade and transformation of the company's business.
Job Description:
The application exploration and implementation of large language models in the fields of e-commerce, gaming, and payments include building multi-language agents with technologies such as function calling, tool usage, RAG, and code interpreters. These technologies are utilized and applied in business areas such as customer service, shopping guidance, video, and search recommendations.
The research and implementation of pre-training and alignment algorithms include ultra-large-scale multilingual pre-training technology, Mixture-of-Experts model training, Instruction Pretraining, SFT, and RLHF. From a multilingual perspective, these efforts aim to reduce the hallucination problem of models, enhance safety capabilities, and improve long-text comprehension and Q&A performance.
Building a service framework and platform for large models, as well as accelerating inference, involves creating online service architectures, data, and evaluation platforms. This also includes exploring and implementing inference acceleration strategies such as Medusa, Speculative Decoding, multi-LoRA inference, and Pruning.
Establishing a comprehensive evaluation system for Southeast Asian multilingual large models involves providing standardized model evaluation capabilities and creating comprehensive evaluation datasets. This system will drive and refine improvements in large language models through model evaluation, addressing issues encountered in practical business scenarios and during technical iteration and optimization processes.
The data mining and optimization of large model algorithms involve constructing data systems for multilingual pre-training, instruction fine-tuning, and human preference behaviors. This includes establishing a comprehensive engineering system for model fine-tuning and dataset preparation to effectively improve the quality and delivery capability of datasets.
Requirements:
Bachelor’s degree or above in Computer Science or related fields.
Excellent coding skills, data structure and basic algorithm skills, proficiency in Python/Pytorch coding, and harness the Hands-on ability.
Familiar with NLP and CV related algorithms and technologies, and those who are familiar with large model training and RL algorithms.
Familiar with the basic principles and training methods of industry-leading LLM (such as GPT, LLaMA), or familiar with the basic principles and training methods of mainstream multi-modal large models (such as Flamingo, LLaVA), and have research experience in text generation or dialogue systems, etc.
Excellent problem analysis and solving skills, able to deeply solve problems in large model training and application.
Good communication and collaboration skills, able to explore new technologies with the team and promote technological progress.
#J-18808-Ljbffr