oumi

Why we included this project

Oumi is a practical pick when your team works with open-weight models and doesn't want to maintain separate toolchains for each stage of the model lifecycle. It handles fine-tuning (SFT, LoRA, DPO, GRPO) on families like Qwen, Gemma, gpt-oss, and DeepSeek-R1, then lets you benchmark, generate synthetic training data, and deploy to vLLM or llama.cpp from the same YAML configuration. That config works on a laptop or a multi-node cluster, which keeps experiments reproducible. It also ships a recipe library and LLM-as-a-judge evaluation helpers, and it builds on vLLM, TRL, and Transformers instead of replacing them, so it drops into existing stacks. For small engineering teams, the cloud job launcher and deploy CLI cover the path from research to production without extra infrastructure glue.