roboflow/notebooks reads as a package but is a folder of tutorials with no pin, no license, and a second org name in every Colab link

Sixty-one notebooks sit behind one autogenerated table

The repository is a directory of Jupyter notebooks, not a distributable package. The index in the README is a section headed model tutorials (61 notebooks) whose rows each carry a file path under notebooks/, a Colab link, a Kaggle link, and pointers to whatever upstream project or paper the model comes from. The filenames are the only reliable statement of scope, and they name tasks directly: notebooks/rf-detr-keypoint-detection.ipynb, notebooks/train-yolo26-object-detection-on-custom-dataset.ipynb, notebooks/train-yolo26-instance-segmentation-on-custom-dataset.ipynb, notebooks/how-to-segment-images-with-segment-anything-3.ipynb, notebooks/how-to-segment-videos-with-segment-anything-3.ipynb, notebooks/open-vocabulary-object-detection-with-qwen3-vl.ipynb, notebooks/how-to-perform-ocr-with-glm-ocr.ipynb, notebooks/how-to-track-objects-with-bytetrack-tracker.ipynb, notebooks/object-detection-with-gemini-3-5-flash.ipynb, and notebooks/how-to-finetune-rf-detr-on-segmentation-dataset.ipynb. The top-level entries are .github/, .gitignore, CONTRIBUTING.md, README.md, assets/, automation/, and notebooks/. Nothing at the top level installs anything.

Colab and Kaggle links are generated from a roboflow-ai path

Every hosted-runtime link in the table is built from a github.com/roboflow-ai/notebooks path, while the repository itself is roboflow/notebooks. Copy the pair for the segmentation notebook exactly as the table writes it:

https://colab.research.google.com/github/roboflow-ai/notebooks/blob/main/notebooks/how-to-segment-images-with-segment-anything-3.ipynb
https://kaggle.com/kernels/welcome?src=https://github.com/roboflow-ai/notebooks/blob/main/notebooks/how-to-segment-images-with-segment-anything-3.ipynb

If a hand-written note or script of yours uses the roboflow path, you have produced a different URL from the one the table advertises, and whether both resolve is not something the table states. The Kaggle link is a kaggle.com/kernels/welcome URL with the notebook path passed as a src parameter.

The table header is a hard warning against editing the index

Above the rows sits an HTML comment that reads WARNING: DO NOT EDIT THIS TABLE MANUALLY. IT IS AUTOMATICALLY GENERATED. and then HEAD OVER TO CONTRIBUTING.MD FOR MORE DETAILS ON HOW TO MAKE CHANGES PROPERLY, with a marker named AUTOGENERATED-NOTEBOOKS-TABLE just above it. The top-level automation/ directory may be where the table is generated, but its contents are not shown. So a fork should not add a row by editing README.md and should go to CONTRIBUTING.md. The column header also advertises open in colab / kaggle / sagemaker studio lab, a third runtime, while the visible rows populate only the first two. Read the header as a schema with a half-filled field.

The 1.0.0 tag predates every notebook the table lists

The only release is 1.0.0, published 2022-12-01, and the last push to the main branch is dated 2026-09-22. That gap is the whole maintenance story: commits keep arriving, tags do not. The consequence is that you cannot pin this collection. If a notebook stops working because a hosted runtime, a public checkpoint, or a dataset URL changed, there is no earlier release to return to, and the 1.0.0 tag identifies no particular set of notebooks. A tutorial collection has good reason to age, but it means the reader is the person who finds the breakage, with no version history to bisect.

Filenames disagree with the greeting about which model version is taught

Three lists of models in this repository do not match. The repository description names ResNet, RF-DETR, YOLO11, SAM 3, and Qwen3-VL. The greeting paragraph names YOLOv11, SAM 2, Florence-2, PaliGemma 2, and Qwen2.5-VL, and names the task set as object detection, segmentation, pose estimation, data extraction and OCR. The table filenames go further still, with YOLO26 in two training notebooks, SAM 3 in the image and video segmentation notebooks, Qwen3-VL in the open vocabulary notebook, a Gemini model for object detection, GLM-OCR for OCR, and ByteTrack for tracking. A reader who skims the greeting to decide what is covered will not learn that YOLO26 notebooks exist or that the Qwen path has moved to Qwen3-VL. Use the filenames, and treat the greeting as an older snapshot of the same list.

First run is a hosted runtime link, not an install command

The top level has no package manifest, so there is nothing to import. Access is one notebook at a time, through the Colab and Kaggle links quoted above. The path is the unit of access rather than a package name, which is why the image segmentation tutorial and the video segmentation tutorial are two separate files that you reach independently. The consequence for a reader is that you never get a snapshot: you either point a hosted runtime at a blob URL, or you take the .ipynb file and own the environment it happens to describe. Neither route gives you a version you can diff against a colleague's, which is the first thing to settle before building on a tutorial here.

What the collection cannot do: record the dependency set you need

This repository cannot tell you which version of a training stack, a transformers release, or a detector package a given notebook expects, because there is no requirements file at the top level. It also gives you no single entry point across the collection: the repository and paper column sends you to different upstreams row by row, including roboflow/rf-detr, roboflow/trackers, ultralytics/ultralytics, facebookresearch/sam3, QwenLM/Qwen3-VL, and zai-org/GLM-OCR. So notebooks that touch rf-detr or trackers depend on those separate projects. If reproducibility matters to you, take the script, freeze the environment yourself, and pin the upstream project the table names.

What the collection cannot answer: licensing and model provenance

Two gaps matter before you build on these tutorials. The first is licensing: this repository's license is unknown and the top level has no LICENSE file, so code you copy out of a notebook arrives with no stated terms attached to it. The second is provenance: the repository and paper column is empty on some rows, and the object detection with Gemini notebook is one of them, with an empty complementary materials column too, which leaves a model name and a file path as the only references. The consequence is that neither column settles what you are allowed to ship, so check this repository's terms and the upstream model's terms as two separate questions before any fine-tuned result reaches a product.

Editorial conclusion

Treat this as a directory of starting scripts rather than a dependency: there is no package to import, no release that identifies a notebook set, and no license stated in the repository, so confirm the terms of this repository and of each upstream model before shipping anything derived from a notebook. If you need a reproducible fine-tuning run, copy the file, freeze its environment, and follow the upstream project named in the table separately, because that upstream, not roboflow/notebooks, is where the breaking changes arrive.

Frequently asked questions

Official sources

1. Issues
2. Project website
3. README
4. Releases
5. roboflow/notebooks on GitHub