# roboflow/maestro

streamline the fine-tuning process for multimodal models: PaliGemma 2, Florence-2, and Qwen2.5-VL

Repository: https://github.com/roboflow/maestro
Canonical: https://ross.abutalabs.com/products/roboflow-maestro
Homepage: https://maestro.roboflow.com
Language: Python
License: Apache-2.0
License Family: permissive
Topics: captioning, fine-tuning, florence-2, multimodal, objectdetection, paligemma, phi-3-vision, transformers, vision-and-language, vqa, qwen2-vl
Last push: 2026-08-24T17:52:24+00:00

## Health v2 (maintenance only)
Score: 62/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 99, release rhythm 8, longevity 72
- inputs: {"age_days": 1013, "days_push": 9, "days_rel": 574, "gap_med": null, "n_releases_24m": 1}
- flags: none
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 2694, forks 222 (observed 2026-08-28T04:07:11.111446+00:00)

## What it is
maestro is a Python library from Roboflow that streamlines fine-tuning of multimodal vision-language models such as Florence-2, PaliGemma 2, and Qwen2.5-VL. It encapsulates configuration, data loading, reproducibility, and training loop setup into ready-to-use recipes with LoRA/QLoRA support.

## Use cases
- fine-tune a vision-language model for object detection
- train Florence-2 on custom captioning data
- fine-tune Qwen2.5-VL to extract JSON from images
- adapt PaliGemma 2 for visual question answering
- run LoRA fine-tuning of VLMs on a free Colab GPU
- fine-tune a multimodal model without writing training loop code

## When to choose
- you want a simple, recipe-driven way to fine-tune supported VLMs like Florence-2, PaliGemma 2, or Qwen2.5-VL
- you need LoRA/QLoRA fine-tuning for vision-language tasks like detection, captioning, VQA, or JSON extraction
- you prefer best-practice defaults over hand-rolling a training pipeline

## When to avoid
- you need to fine-tune models outside the supported list (Florence-2, PaliGemma 2, Qwen2.5-VL, Phi-3-vision)
- you require full control over every aspect of the training loop or custom architectures
- you need text-only LLM fine-tuning with no vision component

## Facets
- artifact type: library
- maturity: active
- function: llm-training, machine-learning, deep-learning, image-processing, cli
- domain: machine-learning, deep-learning, computer-vision, artificial-intelligence, large-language-models
- platform: python, cli
- tags: fine-tuning, vision-language-models, lora, qlora, paligemma, florence-2, qwen2-vl, multimodal, object-detection, vqa, image-captioning, transformers, gpu

## Member repositories
- roboflow/maestro (main) score 62

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:07:11.111446+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T02:16:21.092856+00:00, confidence not recorded.
  - readme: https://github.com/roboflow/maestro (fetched 2026-08-28T04:07:11.111446+00:00, sha 69d87507f899)
  - homepage: https://maestro.roboflow.com (fetched 2026-08-29T09:59:30.460821+00:00, sha 1e06227ac77e)
- Data as of 2026-08-30T08:39:29.467469+00:00.