Researcher - Post-Training, Reinforcement Learning & Training Systems

Preferred

• Experience with CUDA or Triton, long-context training, sparse models, asynchronous RL, stateful agent environments, distillation, or deployment-aware post-training would be especially valuable.

You will

Own the end-to-end post-training pipeline from pretrained checkpoint to production candidate. • Design and execute full-parameter and parameter-efficient SFT. • Implement preference optimisation, RLHF, RLAIF, reinforcement learning with verifiable rewards, and related methods. • Develop training strategies for reasoning, coding, tool use, multilingual behaviour, and long-horizon agent tasks. • Integrate reward models, verifiers, critics, graders, and process- or outcome-based rewards. • Build scalable rollout-generation systems for iterative and on-policy training. • Design multi-stage curricula combining SFT, reinforcement learning, rejection sampling, distillation, and policy consolidation. • Scale training across multiple machines and accelerators using appropriate combinations of data, tensor, pipeline, sequence, context, or expert parallelism. • Select sharding, precision, checkpointing, optimiser, batch-size, sequence-length, and activation-recomputation strategies. • Estimate memory, communication, throughput, rollout capacity, and compute requirements before launching major runs. • Profile and improve accelerator utilisation, communication efficiency, data loading, and end-to-end training time. • Diagnose numerical instability, communication failures, out-of-memory errors, stragglers, checkpoint issues, and convergence regressions. • Investigate reward hacking, entropy collapse, KL drift, stale rollouts, mode collapse, grader exploitation, and benchmark overfitting. • Build reliable checkpointing, recovery, monitoring, and reproducibility procedures. • Collaborate closely with data, evaluation, infrastructure, and inference teams. • Contribute clean, tested code, technical reports, and operational runbooks.

We are looking for demonstrated experience in most of the following areas: • Ownership of large-scale language-model training or post-training runs across multiple machines and accelerators. • Experience with workloads for which straightforward single-node training or pure data parallelism was insufficient.. • Deep proficiency with Python, PyTorch, autograd, mixed precision, optimisation, and distributed execution. • Practical experience with PyTorch Distributed, FSDP, DeepSpeed, Megatron-Core, or an equivalent framework. • Ability to select parallelism and sharding strategies based on model, sequence, memory, and network constraints. • Strong understanding of SFT, preference optimisation, reinforcement learning, reward modelling, KL regularisation, sampling, and training stability. • Experience operating high-throughput inference or rollout systems as part of a training loop. • Ability to debug across model code, distributed communication, numerical optimisation, data, and infrastructure. • Strong experimental design and the ability to distinguish algorithmic improvements from evaluation or systems artefacts. • Experience building reliable, observable, and reproducible research software. • Personal ownership of consequential decisions affecting a substantial training programme. A PhD is not required. We value exceptional technical work, strong judgement, and demonstrated ownership. Relevant stack • Python and PyTorch. • PyTorch Distributed and FSDP. • DeepSpeed, Megatron-Core, or comparable frameworks. • Hugging Face Transformers. • CUDA and NCCL. • vLLM, SGLang, or similar rollout engines. • Ray, Slurm, Kubernetes, or comparable orchestration systems. • MLflow or Weights & Biases. • Docker, GCP, GitLab CI, profiling, monitoring, and pytest.

You may be a strong fit if you: • Enjoy working at the intersection of model research and distributed systems. • Can move from paper reproduction to reliable scaled implementation. • Are comfortable taking responsibility for expensive and operationally demanding experiments. • Approach failures methodically across algorithms, data, numerical stability, and infrastructure. • Care about held-out capability and reliability, not only training loss or reward. • Want meaningful ownership of a complete model programme. Location and work style We offer full-time employment in Switzerland. • Remote work is supported. • The team gathers approximately one week per month in a Swiss office. • Exceptional candidates elsewhere in Europe may be considered..