maestro
A tool from Roboflow that streamlines fine-tuning multimodal vision-language models such as Florence-2, PaliGemma 2, and Qwen2.5-VL.
maestro is a tool from Roboflow that streamlines fine-tuning multimodal vision-language models, handling configuration, data loading, reproducibility, and the training loop so a user does not have to assemble those pieces by hand. It ships ready-to-use recipes for popular models including Florence-2, PaliGemma 2, and Qwen2.5-VL, covering tasks such as object detection and JSON data extraction.
Its 1.0.0 release added support for all three model families with LoRA, QLoRA, and graph-freezing options to keep hardware requirements manageable, a single CLI and SDK to reduce code complexity, and a consistent JSONL data format across models. Training can be started from the command line by specifying the dataset location, epochs, batch size, optimization strategy, and metrics, or from Python by calling a model-specific train function with a configuration dictionary, and example notebooks let a model be fine-tuned for free in a hosted notebook.
It is written in Python and released under the Apache License 2.0.