Finetuning

Supports SFT, DPO, RLVR, and RLAIF techniques with proper scaffolding for each. The workflow is thorough, maybe overly so: it checks prerequisites like use case specs and model selection, handles EULA acceptance for Meta models, and even helps you build custom RLVR reward functions if needed. Code generation follows strict templates with exact imports and minimal improvisation. Worth noting that it adapts output based on model family (strips EULA code for non-Meta, removes certain hyperparameters for Nova). If you're doing one-off experiments this might feel heavy, but for repeatable training pipelines it keeps you from missing steps.