Model Training jobs

These are the roles that actually run big training jobs: pretraining and continued-pretraining at labs, serious fine-tuning everywhere else. The craft is data mixtures, hyperparameters, and distributed-training debugging — knowing what to check when loss spikes at step 400,000. Rarer than a job board makes it look, because many “training” postings are really inference work in costume.

FSDP, DeepSpeed, and Megatron variants are the standard vocabulary, with JAX at some labs and raw PyTorch elsewhere. LoRA and quantization-aware training cover the fine-tuning end; NCCL debugging stories cover the rest. Compensation runs high, and so does the intensity when a run costs six figures a week.