maestro

Why we included this project

Fine-tuning a large vision-language model usually means fighting training loops, dataset formats, and dependency conflicts before you get any real results. Maestro packages that routine for Florence-2, PaliGemma 2, and Qwen2.5-VL into one CLI and Python API: you point it at a dataset, set epochs, batch size, and optimization strategy, and it handles the rest. LoRA and QLoRA support keep hardware demands modest, which matters for small teams on limited GPUs or free Colab notebooks, and the repo ships runnable recipes for each model. A config-driven setup with a consistent JSONL format keeps runs reproducible, and the same tooling covers object detection, captioning, and JSON extraction work. If you want to adapt these models without assembling the training plumbing yourself, this is a practical way in.