gemma/diffusion/hackable_diffusion_adapter/configs/README.md

Kauldron Configurations for SFT Training

This directory contains the Kauldron configuration files defining the models, datasets, training losses, optimizers, schedules, and evaluators for SFT training.

The configurations are flat, self-contained files defining the end-to-end setup (model, dataset, training losses, optimizer, schedule, and evaluators) for each task.

Configuration Files

• sft_sudoku.py (Sudoku):

  • Configures SFT training with LoRA on the sudoku dataset.

• sft_sudoku_full.py (Sudoku):

  • Configures full weight SFT training on the sudoku dataset.

• sft_pubmedqa.py (PubMedQA):

  • Configures Parameter-Efficient SFT training using LoRA on PubMedQA.