Researchers have fine-tuned the MedGemma 4B and 27B models for oncology tasks using Unsloth, completing the process on a single DGX Spark. The resulting OncoLLM models, along with the training dataset and a detailed report, are publicly available.

  • The 4B model was fine-tuned using LoRA, while the 27B model utilized QLoRA.
  • A technical note highlights that Unsloth version 2026.9.12 scales the 4B adapter by approximately 0.5 instead of the nominal alpha/r ratio of 0.25.
  • The source models and code are credited to Grujowmi (ANTOINE Pierre) on Hugging Face, with the full report hosted on Zenodo.