百度公有云大模型训练加速实习生（J107176）
Core
Design post-training alignment strategies for multimodal and text-to-video scenarios to enhance model performance, stability, and generalization.
Role type
Intern, Large Model Training Acceleration
Builds
Production-ready large language models optimized for specific downstream tasks
Domain
Artificial Intelligence, Large Language Models, Distributed Systems
Deliverable
production ML models
Required skills
Large model training, Distributed systems, High-performance computing, SFT, DPO, RLHF, GRPO, Knowledge distillation, PyTorch, Transformers, DeepSpeed, Megatron-LM, FSDP
Preferred skills
Billion-parameter model training, Multimodal model training, Video generation model training, MoE training, Long context training, Sparse model training, Large-scale reinforcement learning, Preference data construction, Reward model training, Online reinforcement learning, Automated evaluation, Safety alignment
Technologies
PyTorch, Transformers, DeepSpeed, Megatron-LM, FSDP
Responsibilities
Design and implement post-training alignment methods (SFT, DPO, RLHF, GRPO, knowledge distillation) for multimodal and video generation tasks; Construct, clean, and analyze training data to optimize training strategies and loss functions; Accelerate distributed training by optimizing compute, memory, communication, and data loading pipelines; Diagnose and resolve training bottlenecks including operator efficiency, communication congestion, load imbalance, and multi-machine stability issues.