Post-Training Engineer - Apertus

EPFL · Lausanne

Contribute to the development, execution, and evaluation of scalable post-training workflows for Apertus.
Build and maintain containerized environments for LLM post-training and RL workloads.
Adapt containers and dependencies for execution on Alps / CSCS infrastructure.
Run and monitor Slurm-based training and evaluation jobs.
Debug failures related to distributed execution, checkpointing, filesystem performance, networking, and GPU utilization.
Help maintain reproducible training recipes, configuration files, launch scripts, and documentation.
Work with researchers and CSCS engineers to improve the reliability and performance of large-scale experiments.
Support SFT, preference optimization, and reinforcement learning workflows.
Build and run RL environments for tasks with verifiable outcomes, such as mathematics, code, tool-use, and reasoning.
Develop reward modeling, reward calibration and verifier-based training.
Generate and validate synthetic or gym training tasks.
Run ablation studies comparing algorithms, reward functions, data mixtures, hyperparameters, and infrastructure settings.

MSc or PhD in Computer Science, Data Science, Artificial Intelligence, Machine Learning, or a related field.
Experience in AI and neural network architectures.
Strong collaboration and communication skills and ability to work across research and engineering teams.

Experience with Slurm or another HPC workload manager.
Experience building or adapting containers for HPC or GPU clusters.
Experience with LLM fine-tuning, post-training, preference optimization, or reinforcement learning.
Familiarity with distributed training concepts such as data parallelism, tensor parallelism, pipeline parallelism, checkpointing, and GPU communication.
Experience with frameworks such as veRL, slime, Megatron-LM, DeepSpeed, TRL, vLLM, SGLang, or similar tools.
Experience with RL for LLMs, online policy optimization, reward modeling, or RLVR.
Experience creating verifiable tasks for mathematics, code, reasoning, or tool use.
Familiarity with lower-level GPU/distributed libraries such as NCCL, Transformer Engine, FlashAttention, or communication backends.
Experience with large-scale evaluation pipelines.

A stimulating academic environment at one of the world's leading technical universities
The opportunity to work with state-of-the-art supercomputing infrastructure and cutting-edge AI research
Collaboration with top researchers and engineers from EPFL, ETH Zürich, CSCS, and other Swiss institutions
Flexible working arrangements, including options for remote work
Professional development opportunities, including conference attendance and specialized training
The chance to contribute to open-source projects with global impact
Access to the broader Swiss academic ecosystem and industry partnerships
Being part of Switzerland's sovereign AI development, working on technology with national significance

Ähnliche Stellen