Rewrite dataset card: gallery, field table, per-type licences

README.md

@@ -1,185 +1,165 @@
 ---
 license: other
 license_name: mixed-per-plugin-licences
-pretty_name: CaptchaOCR
-task_categories:
-- image-to-text
-language:
-- en
-- ru
-tags:
-- captcha
-- ocr
-- minecraft
-- anti-bot
-- synthetic
-- parquet
-modality: image
+license_link: https://huggingface.co/datasets/Titlehhhh/captchaocr-dataset/blob/main/LICENSE
+pretty_name: CaptchaOCR — Minecraft anti-bot captchas
+task_categories: [image-to-text]
+language: [en, ru]
+tags: [captcha, ocr, minecraft, anti-bot, synthetic, parquet, image]
 size_categories: 100K<n<1M
 configs:
 - config_name: default
   data_files: data/*.parquet
 ---
 
-# CaptchaOCR — Minecraft anti-bot captchas as Parquet
+<img src="https://huggingface.co/datasets/Titlehhhh/captchaocr-dataset/resolve/main/assets/hero.png" alt="Six captcha styles from Minecraft anti-bot plugins: map codes, a chat code, Cyrillic symbols and a web captcha" width="100%">
 
-**20 captcha types · 200,000 labelled samples · 9 Parquet shards ·
-1.9 GiB of PNG bytes · 1.9 GiB total**
+# CaptchaOCR — Minecraft anti-bot captchas as training data
 
-This repository holds the **CaptchaOCR** dataset: captchas drawn by open-source Minecraft
-anti-bot plugins, re-rendered by this project's **own Java 21 generators**, one row per
-image, with every label field preserved byte-for-byte in Parquet.
+Minecraft anti-bot captchas, re-rendered by this project's Java 21 generators and packed one image per Parquet row: image, answer, drawn text and answer band.
 
-The previous release stored loose `PNG + labels.jsonl` files under `data/out/` and
-`data/out_ru/`.  That tree has been replaced by these shards; the label values are unchanged.
+![samples](https://img.shields.io/badge/samples-200%2C000-blue)
+![types](https://img.shields.io/badge/captcha_types-20-blueviolet)
+![plugins](https://img.shields.io/badge/upstream_plugins-13-informational)
+![size](https://img.shields.io/badge/size-1.9_GiB-lightgrey)
+![format](https://img.shields.io/badge/format-Parquet-150458)
+![task](https://img.shields.io/badge/task-image--to--text-success)
+![license](https://img.shields.io/badge/license-mixed_per_plugin-orange)
 
-## Files
+## What's inside
 
-| shard | rows | bytes |
+- **20 captcha types** from **13 upstream names** — 6 pixel-faithful ports, 10 clean-room reimplementations, 4 reconstructions with no upstream plugin.
+- **200,000 images**, 10,000 per type, one Parquet row each, with every label field from `labels.jsonl` in the same row.
+- **9 Parquet shards, 1.9 GiB** — 30–248 MiB each, no loading script, no remote code.
+- **`roi` in 12 of 20 types** — 119,984 answer bands; the other 80,016 rows need the whole frame.
+
+## Gallery
+
+Six examples of these types, answer under each image.
+
+| <img src="https://huggingface.co/datasets/Titlehhhh/captchaocr-dataset/resolve/main/assets/gallery/01.png" width="170" alt="Krypton map captcha with a five-letter code"><br><sub>`Krypton/map` — `Wu43b`</sub> | <img src="https://huggingface.co/datasets/Titlehhhh/captchaocr-dataset/resolve/main/assets/gallery/02.png" width="170" alt="mapcha map captcha with a four-letter code"><br><sub>`mapcha/map` — `ly8r`</sub> | <img src="https://huggingface.co/datasets/Titlehhhh/captchaocr-dataset/resolve/main/assets/gallery/03.png" width="170" alt="XProtect chat code captcha"><br><sub>`XProtect/code` — `5dqnZ9`</sub> |
 |---|---|---|
-| `train-00000-of-00009.parquet` | 51,560 | 238,306,789 |
-| `train-00001-of-00009.parquet` | 5,999 | 259,993,245 |
-| `train-00002-of-00009.parquet` | 5,939 | 259,990,006 |
-| `train-00003-of-00009.parquet` | 5,898 | 260,015,557 |
-| `train-00004-of-00009.parquet` | 28,034 | 249,580,369 |
-| `train-00005-of-00009.parquet` | 50,704 | 238,254,361 |
-| `train-00006-of-00009.parquet` | 12,807 | 257,651,098 |
-| `train-00007-of-00009.parquet` | 29,188 | 251,934,144 |
-| `train-00008-of-00009.parquet` | 9,871 | 32,031,576 |
-
-The shards are ordinary Parquet files: one row per sample, the raw PNG bytes in `image`,
-and all label columns alongside.  A shard is about 250 MiB; the row order is deterministic
-(tree, then type, then `labels.jsonl` order).
-
-## Schema
-
-| column | Arrow type | nullable | meaning |
-|---|---|---|---|
-| `image` | `binary` | no | the raw PNG file bytes (decode with PIL / OpenCV) |
-| `file` | `string` | no | PNG file name inside its type directory (`001.png` …) |
-| `type_id` | `string` | no | type path, e.g. `out/Archy-X/Krypton/map` or `out_ru/LimboFilter/symbols` |
-| `plugin` | `string` | no | upstream plugin the type came from |
-| `captcha` | `string` | no | captcha type inside that plugin |
-| `answer` | `string` | no | what the player has to type into the answer field |
-| `text` | `string` | **yes** | what is actually **drawn** on the picture; may be `null` |
-| `prompt` | `string` | **yes** | the task text as the plugin shows it; may be `null` |
-| `roi` | `list<int32>` | **yes** | answer band `[x, y, w, h]`, or `null` if the band is not in the frame |
-| `width` | `int32` | no | PNG width in pixels |
-| `height` | `int32` | no | PNG height in pixels |
-| `source` | `string` | no | where the type came from: upstream class and method (`file#method`) |
-| `tree` | `string` | no | `out` or `out_ru` |
-
-## How to load it
-
-With 🤗 `datasets` (it reads the Parquet shards directly; no loading script, no remote code):
+| <img src="https://huggingface.co/datasets/Titlehhhh/captchaocr-dataset/resolve/main/assets/gallery/04.png" width="170" alt="LimboFilter Cyrillic symbol captcha"><br><sub>`LimboFilter/symbols` — `БЬ5`</sub> | <img src="https://huggingface.co/datasets/Titlehhhh/captchaocr-dataset/resolve/main/assets/gallery/05.png" width="170" alt="Reconstructed web captcha, 150 by 90 pixels"><br><sub>`GalleryCaptcha/ru-web` — `cVejm`</sub> | <img src="https://huggingface.co/datasets/Titlehhhh/captchaocr-dataset/resolve/main/assets/gallery/06.png" width="170" alt="CaptchaVerifyZ map captcha in capital letters"><br><sub>`CaptchaVerifyZ/map` — `AYKZB`</sub> |
+
+## Quickstart
 
 ```python
 import io
-
 from datasets import load_dataset
 from PIL import Image
 
 ds = load_dataset("Titlehhhh/captchaocr-dataset", split="train")
-print(ds)                     # rows: 200,000
 row = ds[0]
-image = Image.open(io.BytesIO(row["image"]))
-band = row["roi"]             # [x, y, w, h] or None
+image = Image.open(io.BytesIO(row["image"])); print(row["answer"], row["roi"])
 ```
 
-Filter by type while loading:
+<details>
+<summary>Filter one type, or read a shard with pyarrow</summary>
 
 ```python
 ds = load_dataset("Titlehhhh/captchaocr-dataset", split="train")
 subset = ds.filter(lambda row: row["type_id"] == "out/haq/mapcha/map")
-```
-
-Without `datasets` — plain `pyarrow` (or pandas):
-
-```python
-import io
-from pathlib import Path
 
 import pyarrow.parquet as pq
-from PIL import Image
 
-table = pq.read_table("data")                      # a directory of Parquet shards
-# or one shard: table = pq.read_table("data/train-00000-of-00000009.parquet")
-rows = table.to_pylist()                           # each row has the columns above
-image = Image.open(io.BytesIO(rows[0]["image"]))
+table = pq.read_table("data/train-00000-of-00009.parquet",
+                      columns=["type_id", "answer"])
+```
 
-# column-projected / filtered read (predicate pushdown, no image decode):
-answers = pq.read_table("data", columns=["type_id", "answer"])
-map_rows = pq.read_table("data", filters=[("type_id", "=", "out/haq/mapcha/map")])
+`train-00000-of-00009.parquet` is the first of 9 shards; the row order is
+deterministic (tree, type, `labels.jsonl` order).
+</details>
 
-# pandas fallback:
-# df = table.to_pandas()
-```
+## Fields
 
-## The band rule: `roi` may be `null`
-
-`roi` is the rectangle `[x, y, w, h]` of the **answer band** inside the image — the crop a
-recogniser should read.  **A band that does not fit into the frame is not a band**, and in
-that case `roi` is `null`.  Never crop those rows: pass the whole image, or skip them.  A
-row with `roi` set can still contain other text (captions, prompts) outside the band, so use
-the rectangle rather than assuming the answer fills the frame.
-
-## `text` may differ from `answer`
-
-* `answer` — what the player types into the answer field.
-* `text` — what is actually drawn on the picture.
-
-For most types they are identical.  In some types (Russian number words, arithmetic, word
-problems) the drawn text differs from the typed answer; converting one to the other is a
-separate solver step, not an OCR task.  **Train a recogniser on `text`; evaluate a solver on
-`answer`.**
-
-## The 20 types
-
-| `type_id` | plugin | captcha | samples | `roi` rows | `text != answer` |
-|---|---|---|---:|---:|---:|
-| `out/Archy-X/Krypton/map` | `Archy-X/Krypton` | `map` | 10,000 | 10,000 | 0 |
-| `out/InstantlyMoist/Captcha/textcaptcha` | `InstantlyMoist/Captcha` | `textcaptcha` | 10,000 | 0 | 0 |
-| `out/JSH32/Minecaptcha/map` | `JSH32/Minecaptcha` | `map` | 10,000 | 10,000 | 0 |
-| `out/LagBug/XProtect/code` | `LagBug/XProtect` | `code` | 10,000 | 0 | 0 |
-| `out/LagBug/XProtect/map` | `LagBug/XProtect` | `map` | 10,000 | 0 | 0 |
-| `out/haq/mapcha/map` | `haq/mapcha` | `map` | 10,000 | 9,984 | 0 |
-| `out/irootdev/AnBot/captcha` | `irootdev/AnBot` | `captcha` | 10,000 | 0 | 0 |
-| `out/jonesdevelopment/sonar/complex` | `jonesdevelopment/sonar` | `complex` | 10,000 | 0 | 0 |
-| `out/jonesdevelopment/sonar/legacy` | `jonesdevelopment/sonar` | `legacy` | 10,000 | 0 | 0 |
-| `out/jonesdevelopment/sonar/standard` | `jonesdevelopment/sonar` | `standard` | 10,000 | 0 | 0 |
-| `out/just-captcha/ProstoCaptcha/map` | `just-captcha/ProstoCaptcha` | `map` | 10,000 | 10,000 | 0 |
-| `out/ringoame196-s-mcPlugin/ReproductionCAPTCHA/map` | `ringoame196-s-mcPlugin/ReproductionCAPTCHA` | `map` | 10,000 | 10,000 | 0 |
-| `out_ru/BotFilter/map` | `BotFilter` | `map` | 10,000 | 0 | 0 |
-| `out_ru/CaptchaVerifyZ/map` | `CaptchaVerifyZ` | `map` | 10,000 | 10,000 | 0 |
-| `out_ru/GalleryCaptcha/ru-numberword` | `GalleryCaptcha` | `ru-numberword` | 10,000 | 10,000 | 10,000 |
-| `out_ru/GalleryCaptcha/ru-numberword-caption` | `GalleryCaptcha` | `ru-numberword-caption` | 10,000 | 10,000 | 10,000 |
-| `out_ru/GalleryCaptcha/ru-symbols` | `GalleryCaptcha` | `ru-symbols` | 10,000 | 10,000 | 0 |
-| `out_ru/GalleryCaptcha/ru-web` | `GalleryCaptcha` | `ru-web` | 10,000 | 10,000 | 0 |
-| `out_ru/LimboFilter/map` | `LimboFilter` | `map` | 10,000 | 10,000 | 10,000 |
-| `out_ru/LimboFilter/symbols` | `LimboFilter` | `symbols` | 10,000 | 10,000 | 0 |
-
-`roi` coverage counts rows whose band is present; `text != answer` counts rows where the
-drawn string and the typed answer differ.
-
-## Provenance and honesty
-
-* **Every image is our own render.**  All 200,000 PNGs were produced by this project's
-  **Java 21 generators** (`dataset/generator` and `dataset/generator_ru`), not by a live
-  server, not by a third-party service, and not by copying anybody else's pictures.
-* The renders reproduce the upstream algorithms (font, palette, noise, layout, even upstream
-  defects), but the pixels are a re-render, not a screenshot of a real session.
-* The generators are **not seeded**: the samples are random.  The images are one particular
-  render (2026-10-24); the labels and generators are the reproducible part.
-* Reference screenshots of the upstream plugins (`real_shots/` in the working tree) are
-  **not** published here: they are other people's images.
-* No picture, page or answer taken from a third-party commercial service is published here.
-* Numbers in this card are computed from `labels.jsonl` by `tools/export_parquet.py` at
-  export time; they are not typed by hand.
-
-## Licence
-
-The data is published under the licence of the upstream plugin each row came from — there is
-no single licence for the whole dataset, hence `license_name: mixed-per-plugin-licences`.
-The repository's `LICENSE` file holds the per-plugin breakdown and what each licence means
-for you.  A user of this dataset must comply with the licence of the plugin each row came
-from.
+| column | type | meaning |
+|---|---|---|
+| `image` | `binary` | raw PNG bytes; decode with `PIL.Image.open(io.BytesIO(...))` |
+| `file` | `string` | file name inside the type directory (`001.png` …) |
+| `type_id` | `string` | type path, e.g. `out/Archy-X/Krypton/map` |
+| `plugin` | `string` | upstream plugin, e.g. `haq/mapcha` |
+| `captcha` | `string` | captcha style inside the plugin, e.g. `map` |
+| `answer` | `string` | what the player types into the answer field |
+| `text` | `string`, nullable | what is drawn on the image |
+| `prompt` | `string`, nullable | task text the plugin shows |
+| `roi` | `list<int32>`, nullable | answer band `[x, y, w, h]` |
+| `width` | `int32` | PNG width in pixels |
+| `height` | `int32` | PNG height in pixels |
+| `source` | `string` | upstream class and method that produced the style |
+| `tree` | `string` | `out` (Latin) or `out_ru` (Cyrillic) |
+
+`text` is what is drawn; `answer` is what the player types. They differ in 3 types, where a solver step converts one to the other — train a recogniser on `text`, evaluate a solver on `answer`. `roi` is the answer band; when the band does not fit the frame, `roi` is `null`, so pass the whole image and never crop those rows.
+
+## Types
+
+20 types from 13 upstream names. `roi` rows count rows with the band present; `text ≠ answer` counts rows where the drawn string differs from the typed answer.
+
+<details>
+<summary>All 20 types</summary>
+
+| type_id | plugin | captcha | samples | roi rows | text ≠ answer | mode | license |
+|---|---|---|---:|---:|---:|---|---|
+| `out/Archy-X/Krypton/map` | `Archy-X/Krypton` | `map` | 10,000 | 10,000 | 0 | clean-room | No licence — all rights reserved |
+| `out/InstantlyMoist/Captcha/textcaptcha` | `InstantlyMoist/Captcha` | `textcaptcha` | 10,000 | 0 | 0 | faithful | MIT |
+| `out/JSH32/Minecaptcha/map` | `JSH32/Minecaptcha` | `map` | 10,000 | 10,000 | 0 | faithful | MIT |
+| `out/LagBug/XProtect/code` | `LagBug/XProtect` | `code` | 10,000 | 0 | 0 | faithful | MIT |
+| `out/LagBug/XProtect/map` | `LagBug/XProtect` | `map` | 10,000 | 0 | 0 | faithful | MIT |
+| `out/haq/mapcha/map` | `haq/mapcha` | `map` | 10,000 | 9,984 | 0 | faithful | MIT |
+| `out/irootdev/AnBot/captcha` | `irootdev/AnBot` | `captcha` | 10,000 | 0 | 0 | clean-room | GPL-3.0 |
+| `out/jonesdevelopment/sonar/complex` | `jonesdevelopment/sonar` | `complex` | 10,000 | 0 | 0 | clean-room | GPL-3.0 |
+| `out/jonesdevelopment/sonar/legacy` | `jonesdevelopment/sonar` | `legacy` | 10,000 | 0 | 0 | clean-room | GPL-3.0 |
+| `out/jonesdevelopment/sonar/standard` | `jonesdevelopment/sonar` | `standard` | 10,000 | 0 | 0 | clean-room | GPL-3.0 |
+| `out/just-captcha/ProstoCaptcha/map` | `just-captcha/ProstoCaptcha` | `map` | 10,000 | 10,000 | 0 | faithful | MIT |
+| `out/ringoame196-s-mcPlugin/ReproductionCAPTCHA/map` | `ringoame196-s-mcPlugin/ReproductionCAPTCHA` | `map` | 10,000 | 10,000 | 0 | clean-room | No licence file — all rights reserved |
+| `out_ru/BotFilter/map` | `BotFilter` | `map` | 10,000 | 0 | 0 | clean-room | BSD-2-Clause, no commercial hosting |
+| `out_ru/CaptchaVerifyZ/map` | `CaptchaVerifyZ` | `map` | 10,000 | 10,000 | 0 | clean-room | All rights reserved (Modrinth) |
+| `out_ru/GalleryCaptcha/ru-numberword` | `GalleryCaptcha` | `ru-numberword` | 10,000 | 10,000 | 10,000 | reconstructed | Not established (see `source`) |
+| `out_ru/GalleryCaptcha/ru-numberword-caption` | `GalleryCaptcha` | `ru-numberword-caption` | 10,000 | 10,000 | 10,000 | reconstructed | Not established (see `source`) |
+| `out_ru/GalleryCaptcha/ru-symbols` | `GalleryCaptcha` | `ru-symbols` | 10,000 | 10,000 | 0 | reconstructed | Not established (see `source`) |
+| `out_ru/GalleryCaptcha/ru-web` | `GalleryCaptcha` | `ru-web` | 10,000 | 10,000 | 0 | reconstructed | Not established (see `source`) |
+| `out_ru/LimboFilter/map` | `LimboFilter` | `map` | 10,000 | 10,000 | 10,000 | clean-room | AGPL-3.0 |
+| `out_ru/LimboFilter/symbols` | `LimboFilter` | `symbols` | 10,000 | 10,000 | 0 | clean-room | AGPL-3.0 |
+
+This snapshot holds 20 of the 21 types in the generator tree. `out_ru/LimboFilter/digits` was rendered after the snapshot and is not published here.
+</details>
+
+## License
+
+Each row carries the license of the upstream plugin it came from, so the dataset as a whole is mixed (`license_name: mixed-per-plugin-licences`). Several upstream licenses forbid commercial hosting or reserve all rights, so check the table before you ship anything. Full text: [LICENSE](https://huggingface.co/datasets/Titlehhhh/captchaocr-dataset/blob/main/LICENSE).
+
+<details>
+<summary>Licenses by upstream plugin</summary>
+
+| plugin | types | samples | license | mode |
+|---|---|---:|---|---|
+| `Archy-X/Krypton` | 1 | 10,000 | No licence — all rights reserved | clean-room |
+| `BotFilter` | 1 | 10,000 | BSD-2-Clause, no commercial hosting | clean-room |
+| `CaptchaVerifyZ` | 1 | 10,000 | All rights reserved (Modrinth) | clean-room |
+| `GalleryCaptcha` | 4 | 40,000 | Not established (see `source`) | reconstructed |
+| `InstantlyMoist/Captcha` | 1 | 10,000 | MIT | faithful |
+| `JSH32/Minecaptcha` | 1 | 10,000 | MIT | faithful |
+| `LagBug/XProtect` | 2 | 20,000 | MIT | faithful |
+| `LimboFilter` | 2 | 20,000 | AGPL-3.0 | clean-room |
+| `haq/mapcha` | 1 | 10,000 | MIT | faithful |
+| `irootdev/AnBot` | 1 | 10,000 | GPL-3.0 | clean-room |
+| `jonesdevelopment/sonar` | 3 | 30,000 | GPL-3.0 | clean-room |
+| `just-captcha/ProstoCaptcha` | 1 | 10,000 | MIT | faithful |
+| `ringoame196-s-mcPlugin/ReproductionCAPTCHA` | 1 | 10,000 | No licence file — all rights reserved | clean-room |
+
+`mode` says how close the pixels are to the upstream plugin: **faithful** is a port of the upstream drawing code, **clean-room** is a reimplementation from the algorithm, **reconstructed** means no upstream plugin was available.
+</details>
+
+## Provenance
+
+- Every image is a re-render by the Java 21 generators in this repository, not a screenshot and not a copy.
+- 6 types are pixel-faithful ports, 10 are clean-room reimplementations, 4 are reconstructed from reference screenshots.
+- The generators are not seeded: this is one random render (2026-10-24). The labels and the generator code are the reproducible part.
+- The upstream reference screenshots that guided the reconstructions are not published; they are other people's images.
+- No image, page or answer from a third-party commercial service is included.
+
+## Limitations
+
+- Synthetic images: fonts, noise and layout follow the plugin algorithms, not a live server, so expect a gap on real screenshots.
+- 8 types have no `roi`; there the answer band is not in the frame, so feed the whole image.
+- 3 types put a solver step between the drawn string and the answer; OCR alone does not solve them.
+- Other anti-bot plugins are not represented: the snapshot stops at the types listed above.