trlX is a framework designed from the ground up for fine-tuning large language models with reinforcement learning. It supports training via either a provided reward function or reward-labeled datasets. The framework supports distributed training across multiple devices and was published at EMNLP 2023.