README.md

metadata

license: other
license_name: personal-use-only
pretty_name: personal-source
task_categories:
  - text-to-image
  - image-to-text
size_categories:
  - 10K<n<100K
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*.parquet
extra_gated_prompt: >-
  This is a private personal image collection with annotations. Access is
  granted manually, for the owner's collaborators only. You must not
  redistribute the images or annotations.
extra_gated_fields:
  Name: text
  Affiliation: text
  Reason for access: text
  I will not redistribute any part of this dataset: checkbox

personal-source (v7)

Personal training set (images and all annotations) used by the Krea2 / Anima personal LoRA and full fine-tuning runs, version v7 (2026-10-09 update). Gated, manual approval. This card contains aggregate statistics only.

• 19,578 images in 90 subset folders, 99 parquet shards (~500 MB, zstd for text columns; image bytes stored as-is), total 48.9 GB.
• Existing images: 17,123 (original pixels from the personal-source archives). Added in v7: 2,455.
• Removed from the set: two artist folders, one teaser collection, one low-quality character image, and 53 images: 31 early-year images of the newly added artist folders (real source year before 2020 or before the per-artist cutoff), 21 betanonbeet images dated before 2025-01-01, and one todo image.
• Update 2026-10-12: artist tags are @-prefixed in tags / tag_groups_json (AI folders excluded) and folder artists were added where missing; affected captions were regenerated and re-translated.
• Verified: row count = 19578, sha256 and md5 of every stored image match, every image decodes (PIL verify() + full load(), width/height match): 19578 / 19578.

Schema (one row per image)

column	type	description
image	struct{bytes,path}	original image file bytes, exactly as stored in the source archive / download (not re-encoded); path is the file name
folder	string	subset folder of the personal dataset (artist / concept / character / collection folder, 90 values)
file_name	string	original file name (stem + extension); folder + file stem is unique
mime_type	string	MIME type detected from the image bytes (PIL)
width / height	int	pixel size of the stored image
file_size	int	byte size of the stored image
md5 / sha256	string	hash of the stored image bytes (sha256 is also the image id used in the training tables)
source	string	where the pixels come from: archive (original personal-source archives, existing images), hub, r2, cdn, x, gelbooru (images added in v7)
new_group	string	selection group of an image added in v7 (null for existing images)
is_new_v7	bool	true = added in the 2026-10-09 update (v7), false = existing before the update
danbooru_id / gelbooru_id / tweet_id	int / int / string	post ids where known
post_url	string	canonical post URL built from the ids above (Danbooru / Gelbooru / X)
source_url	string	original source URL recorded by the booru post (pixiv / twitter / ...), where known
source_date / source_date_method	string	real source date (YYYY-MM-DD) and how it was derived (booru API, tweet snowflake, pixiv CDN path, file name, ...); null when only an imputed year exists
year / year_source	int / string	annotation year used as the year NNNN prompt token in training, and the rule that produced it
rating	string	g / s / q / e where known (booru rating or PixAI-derived)
tags	string	tag string as used in training, format `prefix
tags_source	string	training_new_meta = the string stored in the v7 training table (new-meta format); composed_from_final_tags = old-format training strings are recomposed into the new format from the final merged tag groups (booru/user tags + PixAI) and the final year / quality / complexity labels
tag_groups_json	string (JSON)	tag groups behind tags (artist, character, copyright, general, rating)
tags_v7_train	string	the raw alternate caption stored in the v7 training table (may still be the old 3-segment format); null when no tag row exists
pixai_status / pixai_tags_json	string / string (JSON)	PixAI tagger output: lists of [tag, score] for general / character / copyright / style / meta / rating
pixai_film_grain / pixai_chromatic_aberration	float	PixAI detector scores
caption_en	string	natural-language caption (English), heretic-Qwen3.8 v7 (OCR text and final labels in the prompt)
caption_zh / caption_ja	string	translations of caption_en (Index-Translate-35B-A3B-preview)
caption_source	string	caption generation run tag
quality_label / quality_score / quality_score_source	string / float / string	final quality label used in training (masterpiece / good / normal / bad / worst) and its score (0-1); source = crystal (production metadata) or personal_run (noobreward + aes25 + aim + popularity pipeline)
complexity_tag / complexity_score	string / float	high visual complexity / low visual complexity (null = middle, not tagged), and the overall complexity score (0-100)
aes25_score / aim_score / noob_p_good	float	individual quality model outputs
booru_score / booru_fav_count	int	booru popularity where known
crystal_matched	bool	image matched a row of the production_0925_crystal Danbooru metadata
ocr_allowed / ocr_status / ocr_has_text / ocr_text	bool / string / bool / string	OCR annotation (VLM). Not run (null) for the add, add-n, todo, todo-n folders; ocr_text null unless status is complete, empty string = no text
ocr_used_in_caption	bool	OCR text was part of the caption prompt
repeats_d01 .. repeats_d04	int	num_repeats of the caption branch in each resolution tier (768 / 1024 / 1280 / 1536 views); the tag branch uses 2x; null where the image has no such tier
train_rows_v7 / val_rows_v7 / split_v7	int / int / string	rows the image contributes to the v7 training / validation tables (incl. repeats) and its split

Load with datasets.load_dataset("SumomoLee/personal-source", split="train") (the image column is decoded as an image).

Statistics

source: archive: 17123, hub: 1026, cdn: 632, r2: 511, x: 273, gelbooru: 13

mime type: image/jpeg: 12784, image/png: 6747, image/webp: 42, image/avif: 5

quality label: normal: 8189, masterpiece: 5470, good: 5105, bad: 496, worst: 318

complexity tag: None: 10780, high visual complexity: 4815, low visual complexity: 3983

year source: danbooru_real_source_rule: 7024, ep13_file_mtime: 3347, crystal_annotation_year: 3281, existing_year_tag: 2500, ep13_tag: 1573, filename_source_date: 1567, tweet_snowflake: 273, gelbooru_real_source_rule: 13

split (v7): train: 19223, validation: 355

OCR status: complete: 17502, None: 2073, error: 3

tags source: composed_from_final_tags: 13470, training_new_meta: 6108

caption source: personal_heretic_v7_20261009: 9427, personal_heretic_v7_20261012: 7204, personal_heretic_v7_20261013_curated: 1961, personal_heretic_v7_20261012_special: 975, personal_heretic_20261008: 11

Years (annotation year): 2026: 6282, 2025: 5792, 2024: 3235, 2023: 1927, 2022: 1081, 2021: 577, 2020: 275, 2019: 214, 2013: 172, 2018: 12, 2011: 6, 2015: 2, 2016: 2, 2010: 1

Image size: width 300-10000 px (mean 1923), height 300-9698 px (mean 2269); file size mean 2.49 MB.

Column fill rates (share of rows with a non-null value)

column	fill
folder	100.0%
source	100.0%
caption_en	100.0%
caption_zh	100.0%
caption_ja	99.9%
tags	100.0%
tags_v7_train	97.9%
pixai_tags_json	99.9%
quality_score	100.0%
complexity_tag	44.9%
complexity_score	100.0%
ocr_text	89.4%
danbooru_id	53.5%
post_url	55.0%
source_url	27.9%
source_date	62.1%
booru_score	52.6%
repeats_d01	96.9%
repeats_d02	98.1%
repeats_d03	90.5%
repeats_d04	89.8%

Notes: ocr_text is non-null for 89.4% of the rows (OCR not run for 4 folders; of the rows where it ran, 7,563 contain text). complexity_tag is null for images in the middle complexity band (not tagged in training). caption_* is missing for 1 image (tag-only row).

Provenance

Built by personal-source-release (2026-10-12) from: the personal-update-20261009 work directory (selection, captions DB, quality / complexity / OCR / PixAI annotation runs, real source dates), the original personal-source archives, the Index-Translate translations of the captions, and the data_t2i_v7rc training table (tags, caption text and repeat information as actually used by the training runs).