AdamCodd/Civitai-8m-prompts

The largest dataset available from Civitai. It contains all prompts extracted from images up to May 5th, 2024, the data was sourced from scratch using my own pipeline.

The dataset underwent these specific preprocessing steps:

1. Only entries with both a populated prompt and a negative prompt were retained.
2. Prompts with less than 3 tags were discarded
3. LoRA/Locon/Lycoris/hypernet were removed
4. Duplicates tags were removed (i.e "1girl, 1girl, 1girl" => "1girl")
5. Cleaning of the trailing commas, newlines, empty words (like 'BREAK'), metadata (Steps, model, seed...) wrongfully included into prompt/negativePrompt fields too.
6. The URL field was transformed into a functional URL (i.e 8ecd1fbd-3075-42e3-9005-926981297a39 changed to https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/8ecd1fbd-3075-42e3-9005-926981297a39/width=512/8ecd1fbd-3075-42e3-9005-926981297a39.jpeg)
7. The nsfw field was recomputed as there were a lot of null, if the field nsfwLevel > 2 then nsfw is set to 1 else 0.
8. Only the following fields were kept: url, nsfwLevel, nsfw, stats, prompt, and negativePrompt (much like AdamCodd/Civitai-2m-prompts).

A raw branch of the dataset is available which contains the unprocessed data, including entries with missing prompts, if you want to process the dataset differently.

Total count:

• 8,655,847 entries for the raw dataset (~40GB unzipped)
• 7,852,309 entries for the processed dataset (~8.7GB unzipped)

Changelog:

05/15

• Changed from nsfwLevel > 1 to nsfwLevel > 2 due to the number of false positives with a strict nsfwLevel > 1.
• The nsfw field is set to 1 or 0 instead of true/false to simplify the preprocessing.

If you want to support me, you can here.