NVIDIA BioNeMo Recipes provides step-by-step workflows for fine-tuning large biological foundation models using LoRA (Low-Rank Adaptation), enabling parameter-efficient adaptation on a single workstation GPU. Two case studies are presented: ESM2-3B (protein language model) fine-tuned for protein secondary structure prediction achieving accuracy competitive with state-of-the-art baselines (Q3: 84.80%, Q8: 74.30%), and Evo2-1B (DNA language model) fine-tuned for splice-site classification, jumping from 52.3% (head-only) to 96.6% accuracy while training only ~1.4% of parameters. The recipes integrate NVIDIA Transformer Engine, Megatron-Bridge, and sequence packing (THD format) for ~5.5x throughput improvement. Both workflows complete in under one hour on an NVIDIA RTX 6000 Blackwell GPU.
