#7373rd of 82 in Training and fine-tuning
maestro
Fine-tuning CLI and Python SDK for vision-language models
- Stars
- 2.7k
- License
- Apache-2.0
- Last commit
- Jul 2026
- Last release
- Feb 2025
- Language
- Python
Missed the latest GitHub check; these facts may be older.
Roboflow is hiring: 38 open roles
Overview
maestro wraps configuration, data loading and the training loop for fine-tuning vision-language models. It ships recipes for Florence-2, PaliGemma 2 and Qwen2.5-VL, with LoRA, QLoRA and graph freezing to limit hardware needs. Training runs through one CLI command or a Python train() function, using a JSONL dataset format.
Who it is for: Engineers fine-tuning vision-language models on custom datasets
Strengths
- Ready recipes for Florence-2, PaliGemma 2 and Qwen2.5-VL
- LoRA and QLoRA support keep hardware requirements down
- Same options via CLI and Python train() function
- Colab notebooks for JSON extraction and object detection
Weaknesses
- Only three model families supported
- Florence-2 and Qwen2.5-VL 7B detection recipes are marked experimental
- Per-model dependencies can clash, so separate Python environments are advised
- Last release 1.0.0 dates from February 2025
What it needs
- Models: Florence-2, PaliGemma 2, Qwen2.5-VL
Also in Training and fine-tuning
See all 82| Rank | Project | Score |
|---|---|---|
| 1 | peftPython library for LoRA and other parameter-efficient fine-tuning of Transformers models | 90 out of 100 |
| 2 | unslothDesktop and web app to run and fine-tune local models | 89 out of 100 |
| 3 | trlPost-training library with SFT, DPO, GRPO and KTO trainers | 89 out of 100 |
| #4 | deepspeedPyTorch library for distributed training of very large models with ZeRO | 87 out of 100 |
| #5 | megatron-lmDistributed training library for large transformer models on NVIDIA GPUs | 87 out of 100 |