Senior Machine Learning Engineer, Model Training and Reinforcement Learning

Sourced from jobapplier.app

Nebius Token Factory is building AI training and post-training capabilities for frontier model improvement. This senior role owns the infrastructure enabling reliable, reproducible, and efficient large-scale training and reinforcement learning experiments. The work spans distributed systems, GPU performance, model training frameworks, RL pipelines, and production engineering.

• Design and run model-training and post-training experiments including SFT, continued pretraining, preference optimization (DPO/IPO/KTO), and RL methods such as RLHF/RLAIF, PPO, and GRPO
• Build reward functions, judge models, verifiers, task environments, and evaluation sets for reasoning, coding, tool use, and agentic workflows
• Create synthetic data and data pipelines including teacher-student generation, self-play, rejection sampling, filtering, and quality scoring
• Build and maintain distributed training and RL infrastructure using Megatron-LM, DeepSpeed, PyTorch FSDP/DTensor, Ray, verl, slime, AReaL, or OpenRLHF
• Implement and debug parallelism strategies and reliable rollout, reward-serving, checkpointing, and experiment-orchestration components
• Profile and improve GPU utilization, memory usage, communication efficiency, training throughput, and inference/serving performance
• Design rigorous evaluations and ablations for capability, instruction following, reasoning, tool use, safety, and regression risk

• Strong Python and PyTorch engineering skills
• Hands-on experience across at least two of: model training, post-training/RL, applied modeling, data pipelines, or large-scale ML systems
• Ability to design rigorous experiments with baselines, ablations, metrics, and failure analysis
• Practical understanding of modern LLM behavior, instruction tuning, preference optimization, and evaluation challenges
• Practical understanding of transformer training bottlenecks, memory pressure, communication overhead, and checkpointing
• Ability to reason quantitatively about model quality, throughput, utilization, reliability, cost, and research velocity
• Strong communication and collaboration skills

• Experience with LLM post-training, RL, agents, reward modeling, synthetic data, or model evaluation
• Experience with RL frameworks such as verl, slime, AReaL, OpenRLHF, TRL, or custom PPO/GRPO/RLHF systems
• Experience with Megatron-LM, DeepSpeed, PyTorch FSDP/DTensor, Ray, Slurm, or Kubernetes on large GPU clusters
• Familiarity with NCCL, CUDA, Triton, Nsight, InfiniBand/RDMA, and H100/H200/B200 clusters
• Publications, open-source contributions, or production impact in LLM post-training, RL, reasoning, coding models, synthetic data, distributed training, or evaluation
• Experience designing agent environments, tool-use tasks, or verifier-based rewards