Fine-tune transformer language models using TRL with support for SFT, DPO, GRPO, and reward model training.