roboflow

maestro: Fine-Tune Vision-Language Models

Maestro packages training workflows for fine-tuning multimodal vision-language models with LoRA and QLoRA. It suits Python developers adapting supported models for tasks such as JSON extraction or object detection, using consistent data and CLI or Python interfaces.

View on GitHub Compare
Stars
2.7k
Forks
225
Stars, last 30 days
+1
Last commit
2 months ago
Contributors
15

Our take

Maintaineda few commits in the last 90 days

Health score

33/100

How it is scored
Activity
31
Community
55
Issues
21
Pull requests
25

+1 stars in the last 30 days

  • Slower to merge pull requests than 100% of the projects we track

Maestro is a focused set of practical fine-tuning workflows, with recent maintenance activity and a broad contributor base, but no release in over a year.

Good fit if

  • You are working in Python and your task and model match one of its supported recipes.
  • You want LoRA or QLoRA options and a shared CLI or API instead of assembling the training setup yourself.
  • Apache-2.0 licensing suits your project, and you can work within the supported Python versions.

Look elsewhere if

  • You need a model or task beyond the listed Florence-2, PaliGemma 2, and Qwen2.5-VL workflows.
  • You need a recently released version or evidence of active issue closure before adopting a training dependency.
  • Your project cannot use Python or needs training workflows beyond the recipes and examples provided.
All health signals
Last commit2026-07-27 (2 months ago)
Commits, last 90 days4
Releases, last 12 months0 (latest 1.0.0, 2025-02-05)
Contributors15 (top contributor: 49% of commits)
Issues closed, last 90 days0
Project age2 years

Checked on 2026-10-07 with the GitHub API.

Project history

Daily GitHub snapshots since Oct 7, 2026. History covers stars and forks; health signals are not tracked historically.

Selected snapshot

Stars

2,696

Forks

226

Change since Nov 21, 2026: 0 stars·0 forks

Overview

Maestro is a Python toolkit for fine-tuning multimodal vision-language models, with ready-to-use workflows for Florence-2, PaliGemma 2, and Qwen2.5-VL. It brings configuration, data loading, reproducibility, and training-loop setup into a shared CLI and Python API, reducing the custom training code needed to adapt these models.

It is aimed at developers and machine-learning teams working with vision-and-language tasks who want an established recipe rather than building a fine-tuning pipeline from scratch. The included workflows cover JSON extraction and object detection, with LoRA or QLoRA options intended to keep hardware demands more manageable.

Key Features

  • Provides model-specific fine-tuning recipes for Florence-2, PaliGemma 2, and Qwen2.5-VL.
  • Supports LoRA and QLoRA optimization strategies, plus graph freezing as described in the project’s release notes.
  • Offers both a command-line interface and a Python training API.
  • Uses a consistent JSONL data format across its workflows.
  • Includes cookbook notebooks for JSON extraction and object detection, including experimental object-detection recipes.
  • Handles configuration, data preparation, reproducibility, and training setup through its core modules.

Use Cases

  • A computer-vision developer can adapt Qwen2.5-VL for object detection using the experimental QLoRA recipe.
  • A data engineering team can fine-tune PaliGemma 2 to extract structured JSON from images.
  • A team evaluating Florence-2 can try its experimental LoRA object-detection workflow before building a custom trainer.
  • Python practitioners can run a supported training configuration from the CLI, or use the API when they need to control it from code.

What you need

Detected in the repository

  • Python >=3.9,<3.13 (from pyproject.toml)
  • A test suite and automated checks on GitHub Actions

License in plain words

Apache-2.0permissive

  • Commercial use: yes
  • Modify and redistribute: yes
  • You must keep: the license, the NOTICE file and a note of your changes
  • Share your changes: no
  • Includes an explicit patent grant from the contributors:

A summary, not legal advice: the LICENSE file is what applies.

Getting Started

Create a model-specific Python environment, then install the relevant extra. For example:

pip install "maestro[paligemma_2]"

See the README for model-specific setup, CLI options, Python usage, and cookbook notebooks.

Considerations

The project has made a few commits recently, but has had no release in over a year, and the measured data shows no issues closed in the last 90 days. Workflows are specific to a short list of models, and the README marks some object-detection recipes as experimental. Training compute needs will depend on the selected model and configuration; LoRA and QLoRA are offered to help manage hardware requirements, but the repository information does not specify a universal hardware baseline. Check each model’s dependency requirements and the project guidance before choosing an environment.

Found this useful?

Share it with someone who would like maestro.

Comparisons