Files
gpu-rent/datasets
Leonid PershinandCursor 6dfbf4e702 Add Civitai Krea2 metadata scrape for train JSONL and Assistent FTS search.
Local civitai-dataset launchers collect ~2000 prompt/params rows without images; search.jsonl is pushed on up for cheap example lookup.

Co-authored-by: Cursor <cursoragent@cursor.com>
2026-08-23 08:13:53 +03:00
..

Civitai Krea2 datasets (local)

Scraped generation metadata for Assistent training + cheap FTS search. No image files.

Not part of gpu-rent Typer CLI and not models.yaml / gpu-rent push. Separate entrypoint:

civitai-dataset.bat          # Windows (or .\civitai-dataset.ps1 / ./civitai-dataset.sh)
# default = all → discover + scrape + split
python -m gpu_rent.civitai_dataset all

Needs CIVITAI_API_TOKEN and preferably CIVITAI_API_HOST=civitai.red in .env.

Commands

Command Flags Role
discover --top-checkpoints (20), --top-loras (30) Popular Krea2 ckpt/LoRAs (+ versions from models.yaml) → catalog/models.json
scrape --target (2000), --per-version, --min-score Gallery metadata only; resume by id; NSFW→rating; skip minor tags
split by_kind/ / by_rating/ / train.jsonl / search.jsonl
all same as discover+scrape discover → scrape → split (launcher default)

Global: --out = app root (writes under datasets/civitai).

Layout

Path Role
civitai/catalog/models.json Discovered Krea2 checkpoints / LoRAs
civitai/catalog/images.jsonl Canonical rows (resume by id)
civitai/by_kind/ checkpoint vs lora
civitai/by_rating/ pg / pg13 / r / x
civitai/train.jsonl Alpaca instruction/output for Assistent Обучение
civitai/search.jsonl Compact FTS index → pushed to VM as Assistent/civitai-examples.jsonl on up / seed-personas

Push of search.jsonl is skipped if the file is missing. Weights stay in models.md (Civitai seed / Models/ / capture).

Git ignores the data; this README stays.

See also: docs/cli.md (civitai-dataset block), docs/extensions.md, docs/local-folders.md.