SlipstreamJobs tracks this role from the company's public career site. Apply directly on the employer's site.
Cosine is building autonomous software-engineering systems for organizations where reliability, security and control matter most. The Lumen Enterprise models are designed to work across real codebases, understanding unfamiliar architectures, making changes across multiple files, using tools, running tests, and producing maintainable software.
You will work across the full training stack, including pre-training, supervised fine-tuning, and reinforcement learning. This is a hands-on role requiring comfort profiling training loops, identifying performance bottlenecks and memory leaks, and implementing new RL algorithms. You'll work closely with research, infrastructure, and product teams to answer critical questions: What should Lumen learn next? Which data and environments teach most effectively? How should we reward correct, maintainable software? How do we train models to work across large repositories and long contexts? How do we measure genuine improvement for engineers?
The models you train will write code, use development tools, run tests, and receive feedback from real software-engineering environments. Your work directly influences how Lumen reasons about code, navigates unfamiliar systems, and improves through experience.
Key responsibilities include: running pre-training, SFT, and RL experiments at scale; optimizing training methods on frontier models; diagnosing and fixing performance bottlenecks in distributed training; training models on long-context, multi-file, and repository-level tasks; developing RL loops using signals from execution, tests, linters, and static analysis; investigating model failure modes; and collaborating with infrastructure, platform, and data teams.
The role emphasizes that the best coding models are not simply those generating plausible code—they must understand context, respect existing architecture, use tools effectively, test their work, and recover from failures. This behavior must be trained, and you'll work directly on the data, environments, rewards, and evaluations that shape Lumen's capabilities.
Cosine operates as an in-office team, five days a week by design, believing the work benefits from close collaboration and shared context. The company values focus, sustainability, and work-life balance over performative overwork.
REQUIREMENTS:
- Strong experience training deep-learning models in a production or research environment
- Deep practical experience with PyTorch or comparable framework
- Experience with distributed training and multi-GPU systems (FSDP, Deepspeed, etc.)
- Strong software-engineering fundamentals with high bar for correctness and maintainability
- Clear communication and ability to explain technical trade-offs
- Experience with large language models, having worked meaningfully on at least one of: pre-training, SFT, preference optimization, RLHF, or RLVR
BONUS QUALIFICATIONS:
- Experience with mixture-of-experts (MoE) architectures
- CUDA experience
- Familiarity with vLLM, SGLang, or other high-performance inference systems
- Experience running jobs on SLURM, Kubernetes, or managed GPU platforms
- Experience with code-model benchmarks (SWE-bench, HumanEval, repository-level evaluations)
- Contributions to open-source model-training, RL, or evaluation tools
