Forge is a training framework that uses YAML configs to manage different model training experiments.

Detail

  • Forge is a YAML-driven training framework, so experiments can be modified by changing configuration files rather than rewriting training code.
  • It is built on the Hugging Face stack, using TransformersHugging Face Trainer, and TRL for training workflows.
    • Transformers = load models, tokenizers, and processors
    • Hugging Face Trainer = run the standard supervised training paths, including SFT and KD
    • TRL = run the current RL-style training path, currently GRPO
  • supports multiple training paradigms:
    • supervised fine-tuning
    • knowledge distillation
    • reinforcement-learning-based post-training, currently via the GRPO path
  • provides experiment infrastructure:
    • W&B logging
    • evaluation
    • checkpointing
    • resume-from-checkpoint
    • SLURM auto-resubmission for long-running jobs