Danbooru dataset cleaned
Tag Cleaning Dataset Update: 430K Corrections and 1.9M Missing Tags Added
As part of my ongoing ViT project for improving anime-image tagging, I’m releasing the current version of my tag-cleaning dataset:
Dataset: https://huggingface.co/datasets/Grio43/Tag_cleaning
This release was produced from a larger dataset containing approximately 6.2 million images. So far, the cleaning process has:
- Corrected approximately 430,000 incorrect tag assignments
- Added approximately 1.9 million missing tag assignments
- Reviewed roughly 9,621 tags
The dataset will continue to receive corrections as the ViT improves. Future review passes will be more targeted, focusing on tags and concepts that remain difficult for the model.
Current dataset expansion
I am also expanding the 6.2 million-image dataset with a small number of supplemental images from carefully selected datasets. These additions target difficult concepts ranging from animals to weapons and other commonly misidentified subjects.
While reviewing the ViT’s failure cases, I found several recurring problems:
- The model sometimes associates a particular object or concept too strongly with a specific character.
- The original Danbooru data contains substantial tagging noise.
- Some visually similar concepts are frequently confused. For example, an image tagged as an animal may actually depict an
animal_girl.
To improve these weaker areas, I plan to add approximately 3% photorealistic images for selected target concepts. This should be small enough to avoid significantly shifting the dataset’s overall distribution, while hopefully helping the model recognize the underlying subject rather than memorizing anime-specific correlations. It may also provide a broader improvement in generalization.
Preview model
The next release will be a low-resolution V1 preview model. All tags will be regenerated from the latest version of the dataset as of August 30, 2026.
The preview model may be somewhat unstable because it will be tuned to create clearer boundaries between closely related tags. Its primary purpose is to help identify and correct additional problems in the dataset before committing the resources required to train a higher-resolution model.
Feel free to use the cleaned data for LoRA training, fine-tuning, tagging experiments, or other research projects. Feedback, especially examples of incorrect or missing tags, is welcome.