Design scalable LLM pipelines using FSDP, DeepSpeed, HuggingFace Accelerate
Lead model development (e.g., LLaMA, Mistral, Phi, Gemma) using LoRA, FlashAttention-2, MoE
Contribute to safety alignment (RLHF, DPO, red-teaming, rejection sampling, calibration)
Support data pipeline audits (tokenizer design, deduplication, privacy, synthetic supervision)
Benchmark across general and medical tasks (lm-eval-harness, HELM, guideline adherence)
Publish and present research at top venues (e.g., NeurIPS, ICLR, ML4H)
Profile
- Proven experience training >1B parameter models with distributed infrastructure
- Expertise in PyTorch, HuggingFace Transformers, DeepSpeed, FSDP
-
Deep understanding of transformer internals and optimization strategies
- Strong Python engineering skills (tests, containers, CI/CD, reproducibility)
- Clear scientific communication and publication experience
Preferred
Familiarity with clinical LLMs or decision support systems
Experience with safety-critical evaluation (e.g., hallucination detection, benchmark leakage)
Contributions to open-source projects
Passion for equity-centered deployment and global health
Our Stack
PyTorch, HuggingFace, DeepSpeed, FSDP, WANDB, Hydra, MLFlow, WebDataset, Slurm, Docker, lm-eval-harness, OpenCompass, MedQA, SwissAlps (CSCS)
Only applications submitted through the online platform are considered. You are asked to supply:
- A brief cover letter (pdf, up to 2 pages).
And in one PDF:
- A CV with a publication list.
- A research statement (pdf, up to 3 pages).
- Contact details for 3 referees.
For any further information, please contact: [XXX]
