Trinity-RFT

Why we included this project

Teams that want to move a model beyond supervised fine-tuning into reinforcement fine-tuning will find a practical starting point here. Trinity-RFT splits the RFT loop into three cooperating pieces: an Explorer that gathers experience through agent-environment interaction, a Trainer that updates weights, and a Buffer that handles data processing. That separation lets you reason about each stage instead of wrestling with one monolithic script. It ships with working examples of GRPO and related algorithms, supports LoRA, multi-modal models, and FSDP2/Megatron parallelism, and even offers a GPU-free backend for experimentation. That makes it useful both for agent developers sharpening a model on domain-specific tasks and for RL researchers prototyping new algorithms on compact, plug-and-play modules, and the tutorials walk through each workflow for teams new to RL-based training.