6DammK9/e621_2024-tags-1ktar

E621 2024 tags only in 1k tar

• Dedicated dataset to align both NebulaeWis/e621-2024-webp-4Mpixel and deepghs/e621_newest-webp-4Mpixel.

• How to use / why I create this: my speedrun to build the dataset

Core logic

• The script building this 1ktar

• The script building the tag from 2 datasets

• Artist and character first. Literally as-is. No fancy "quality" score for pretraining focused dataset.

df3 = pd.DataFrame(data={
    'id': df11["id"].to_list() + df21["id"].to_list(), 
    'rating': df11["rating"].to_list() + df21["rating"].to_list(), 
    'tag_string': [x.replace(" ",", ") for x in df11["tag_string"].to_list()] + [", ".join(x) for x in df21["tags"].to_list()], 
})

How to build the "dataset" with speed

• Get at least 4TB of storage, and around 75GB of RAM. Always make a venv / conda environment for each task.

• (Optional) Download this directly: posts-2024-04-07.parquet and table.parquet

• Download all 1k tarfile with webp via dl-e621-hfhub-nw.py and dl-e621-hfhub-dgs.py

• Rerun that script for this repo (another 1k tarfile).

• Run extract-e621-parallel.py to extract all tars into a single directory.

> python extract-e621-parallel.py
100%|██████████████████████████████████████| 1000/1000 [2:30:51<00:00,  9.05s/it]
Extracted: 1000 iters
Delta: 744438 files

PS H:\e621_newest-webp-4Mpixel> node
Welcome to Node.js v20.15.0.
Type ".help" for more information.
> const fs = require('fs');
> console.log(fs.readdirSync("./kohyas_finetune").length);
8883320

• (Done?) Finally, instead the official guide (a bit messy), follow this reddit post to make the metadata JSON file (with ARB) and start finetuning.