Titlehhhh commited on
Commit
dab1d48
Β·
verified Β·
1 Parent(s): f63f2cb

CaptchaOCR model card: add captcha gallery from the dataset repo

Browse files
Files changed (1) hide show
  1. README.md +198 -186
README.md CHANGED
@@ -1,186 +1,198 @@
1
- ---
2
- license: apache-2.0
3
- library_name: transformers
4
- base_model: PaddlePaddle/PP-OCRv6_medium_rec_safetensors
5
- tags:
6
- - captcha
7
- - ocr
8
- - minecraft
9
- - anti-bot
10
- - pp-ocr
11
- - image-to-text
12
- - pytorch
13
- pipeline_tag: image-to-text
14
- ---
15
-
16
- # CaptchaOCR PP-OCRv6 β€” a text recogniser for Minecraft anti-bot captchas
17
-
18
- > ## Status: the weights are **not published yet**. This repository holds the model card only.
19
- >
20
- > The fine-tune on the [CaptchaOCR dataset](https://huggingface.co/datasets/Titlehhhh/captchaocr-dataset)
21
- > has **not been run to completion yet**, so there is **no checkpoint, no `config.json` and no
22
- > measured metric** in this repository. Nothing here is invented and nothing is extrapolated from a
23
- > probe run: the numbers quoted below are the **measured numbers of the base model before
24
- > fine-tuning**, and they are labelled as such.
25
- >
26
- > **RU:** Ρ€Π΅ΠΏΠΎΠ·ΠΈΡ‚ΠΎΡ€ΠΈΠΉ ΠΌΠΎΠ΄Π΅Π»ΠΈ ΠΏΠΎΠΊΠ° **пуст ΠΏΠΎ вСсам** β€” здСсь Ρ‚ΠΎΠ»ΡŒΠΊΠΎ ΠΊΠ°Ρ€Ρ‚ΠΎΡ‡ΠΊΠ°. ΠžΠ±ΡƒΡ‡Π΅Π½ΠΈΠ΅ Π½Π°
27
- > датасСтС `Titlehhhh/captchaocr-dataset` Π΅Ρ‰Ρ‘ Π½Π΅ Π·Π°Π²Π΅Ρ€ΡˆΠ΅Π½ΠΎ, поэтому **ΠΌΠ΅Ρ‚Ρ€ΠΈΠΊΠΈ появятся послС
28
- > ΠΏΡ€ΠΎΠ³ΠΎΠ½Π°** (плСйслолдСр Π½ΠΈΠΆΠ΅ β€” это Π½Π΅ Π²Ρ‹Π΄ΡƒΠΌΠ°Π½Π½Ρ‹Π΅ числа). Π’Ρ‹ΠΊΠ»Π°Π΄ΠΊΠ° ΠΏΡ€ΠΎΠ±Π½ΠΎΠ³ΠΎ Ρ‡Π΅ΠΊΠΏΠΎΠΈΠ½Ρ‚Π° (2 эпохи)
29
- > ΡΠΎΠ·Π½Π°Ρ‚Π΅Π»ΡŒΠ½ΠΎ Π½Π΅ сдСлана: ΠΎΠ½ Ρ…ΡƒΠΆΠ΅ Π±Π°Π·ΠΎΠ²ΠΎΠΉ ΠΌΠΎΠ΄Π΅Π»ΠΈ.
30
-
31
- ## What this model is
32
-
33
- A fine-tune of **PP-OCRv6 medium** (PaddleOCR's text recogniser) on the CaptchaOCR dataset: the
34
- captchas that open-source Minecraft anti-bot plugins draw, rendered by this project's own Java
35
- generators (46 types, 4600 labelled PNG).
36
-
37
- * **Architecture:** PP-OCRv6 medium recogniser β€” PP-LCNetV4 backbone + CTC head,
38
- `transformers` class `PPOCRV6SmallRecForTextRecognition`
39
- (`model_type: pp_ocrv6_small_rec`), 19.1 M parameters, hidden size 192.
40
- * **Dictionary:** the base model's own 18 710 classes (index 0 = CTC blank, space at the end).
41
- * **Input:** the captcha **text band** (the strip of pixels that carries the answer), not the whole
42
- frame: resized to height 48, width scaled proportionally and right-padded to 320, BGR channel
43
- order, normalised as `(x / 255 - 0.5) / 0.5`.
44
- * **Output:** per-timestep class probabilities; decoded greedily (CTC: collapse repeats, drop the
45
- blank). Greedy decoding is the right decoder here β€” the answers are random strings, so a language
46
- model has nothing to exploit.
47
-
48
- **The band is the whole game.** On the 29 dataset types that carry a band, the base model scores
49
- `acc_micro = 0.5517` on the band and `acc_micro = 0.0345` on the whole frame β€” about 16Γ— worse on
50
- the frame. A solver must crop first.
51
-
52
- ## Provenance and licence
53
-
54
- * **Base model:** [`PaddlePaddle/PP-OCRv6_medium_rec_safetensors`](https://huggingface.co/PaddlePaddle/PP-OCRv6_medium_rec_safetensors)
55
- β€” Apache-2.0, by PaddlePaddle.
56
- * **This model:** that base, fine-tuned on the CaptchaOCR dataset published here:
57
- [`Titlehhhh/captchaocr-dataset`](https://huggingface.co/datasets/Titlehhhh/captchaocr-dataset)
58
- (46 captcha types Β· 4600 labelled PNG Β· 23 upstream plugins).
59
- * **Licence of the weights:** Apache-2.0, inherited from the base model. The fine-tuning code and
60
- this card are released under the same terms.
61
- * **Licence of the dataset:** *not* Apache-2.0 β€” the dataset is split by the licence of the
62
- upstream plugin of each captcha type. The dataset card has the per-type table.
63
-
64
- ## Measured results
65
-
66
- ### This repository's weights
67
-
68
- **No metrics yet β€” the training run has not been completed.** This section will hold exact match
69
- and CER per type and on the type-held-out split when the fine-tune is done. Until then, treat the
70
- numbers below as the **starting point**, not as the result of this repository.
71
-
72
- ### The base model, measured before fine-tuning (baseline)
73
-
74
- These are measured numbers of `PaddlePaddle/PP-OCRv6_medium_rec_safetensors` **as it is**, on the
75
- CaptchaOCR dataset β€” i.e. the point the fine-tune starts from:
76
-
77
- | measurement | value | source |
78
- |---|---|---|
79
- | exact match (micro) on the 7 type-held-out types, band crop | 0.5586 | `train/run_ppocr/` zero-step baseline |
80
- | exact match (median over those 7 types) | 0.60 | `train/run_ppocr/` zero-step baseline |
81
- | CER (micro) on the same split | 0.2561 | `train/run_ppocr/` zero-step baseline |
82
- | exact match (micro) on the 29 band-carrying types, band crop | 0.5517 | `reports/bench_models/ppocrv6-medium.json` |
83
- | exact match (micro) on the same 29 types, whole frame | 0.0345 | `reports/bench_models/ppocrv6-medium.json` |
84
-
85
- Both sources are files in the **CapthaModel** project repository; nothing above is a projection.
86
-
87
- **Why the weights are not here yet.** A short probe run (2 epochs) of the fine-tune was measured
88
- on the same held-out types and it came out **worse than the base model** β€” median exact match
89
- 0.41 after epoch 1 and 0.24 after epoch 3, against 0.60 for the untouched base model. That
90
- checkpoint is deliberately **not** published: a two-epoch probe is not a model, and shipping it as
91
- one would misrepresent it.
92
-
93
- ## How to use it
94
-
95
- > The snippet below is the **preprocessing the project itself uses** (`train/ppocr_data.py`), and it
96
- > was verified end-to-end against the base model on a real sample of the dataset (it decoded the
97
- > held-out sample `JSH32/Minecaptcha/map/001.png`, band `[25, 77, 79, 35]`, as `PTAY`, which is the
98
- > gold answer). It is written for the moment the weights are published β€” until then
99
- > `from_pretrained("Titlehhhh/captchaocr-ppocrv6")` has nothing to load.
100
-
101
- ```python
102
- import json
103
-
104
- import numpy as np
105
- import torch
106
- from huggingface_hub import hf_hub_download
107
- from PIL import Image
108
- from transformers import PPOCRV6SmallRecForTextRecognition
109
-
110
- REPO_ID = "Titlehhhh/captchaocr-ppocrv6"
111
-
112
- model = PPOCRV6SmallRecForTextRecognition.from_pretrained(REPO_ID).eval()
113
- # the dictionary comes with the model repository (preprocessor_config.json)
114
- character_list = json.load(
115
- open(hf_hub_download(REPO_ID, "preprocessor_config.json"), encoding="utf-8")
116
- )["character_list"]
117
-
118
- HEIGHT, MAX_WIDTH = 48, 320
119
-
120
-
121
- def prepare(band: Image.Image) -> torch.Tensor:
122
- """band -> (1, 3, 48, 320) float tensor."""
123
- scale = HEIGHT / max(1, band.height)
124
- w = max(1, min(MAX_WIDTH, int(round(band.width * scale))))
125
- small = band.convert("RGB").resize((w, HEIGHT), Image.BILINEAR)
126
- a = np.asarray(small, dtype=np.float32) / 255.0
127
- a = a[:, :, ::-1] # BGR, as PP-OCRv6 expects
128
- a = (a - 0.5) / 0.5 # paddle normalisation
129
- canvas = np.zeros((3, HEIGHT, MAX_WIDTH), dtype=np.float32)
130
- canvas[:, :, :w] = a.transpose(2, 0, 1)
131
- return torch.from_numpy(canvas)[None]
132
-
133
-
134
- def greedy_decode(probs) -> str:
135
- """CTC greedy decode: collapse repeats, drop the blank."""
136
- best = probs.argmax(dim=-1)[0].tolist()
137
- chars, prev = [], -1
138
- for k in best:
139
- if k != prev and k != 0:
140
- chars.append(character_list[k])
141
- prev = k
142
- return "".join(chars)
143
-
144
-
145
- image = Image.open("captcha.png")
146
- x, y, w, h = (25, 77, 79, 35) # the answer band: take it from labels.jsonl
147
- band = image.crop((x, y, x + w, y + h))
148
-
149
- with torch.no_grad():
150
- probs = model(pixel_values=prepare(band)).last_hidden_state
151
-
152
- print(greedy_decode(probs))
153
- ```
154
-
155
- The band (`roi`) is not optional: on scene-like captchas the answer is one line among several, and
156
- the whole frame makes the model read the wrong line.
157
-
158
- ## Reproducing
159
-
160
- The training code lives in the **CapthaModel** repository:
161
-
162
- ```powershell
163
- .venv-gpu\Scripts\python.exe train\finetune_ppocr.py --epochs 30 --batch-size 32
164
- ```
165
-
166
- `train/finetune_ppocr.py` fine-tunes `PaddlePaddle/PP-OCRv6_medium_rec_safetensors` on the
167
- dataset, holds out whole **plugins** (never random splits β€” a random split leaks the exact gold
168
- answer of 34 % of the test samples), and reports exact match and CER per type plus the median over
169
- types. `train/ppocr_data.py` holds the dataset loader and the preprocessing shown above.
170
-
171
- ## Limitations
172
-
173
- * **Only the text band is read.** This is a recogniser, not a solver: it does not reverse strings,
174
- does not turn Russian number words into digits and does not compute arithmetic.
175
- * **No Cyrillic in the dictionary.** The base model's 18 710 classes contain no Cyrillic at all, so
176
- the Cyrillic captcha types of the dataset cannot be expressed in this model's alphabet.
177
- * **The data is synthetic.** Every image is a render of this project's own generators β€” faithful to
178
- the upstream algorithms, but not a screenshot of a live server.
179
- * **Generalisation is the hard part.** A recogniser can memorise a plugin's style instead of
180
- learning the characters; only the type-held-out numbers mean anything.
181
-
182
- ## Citation
183
-
184
- This model is a fine-tune of PP-OCRv6 by PaddlePaddle (Apache-2.0). The dataset is the
185
- CaptchaOCR dataset, re-rendered from open-source Minecraft anti-bot plugins. See the dataset card
186
- for the provenance and the per-type licence of every captcha type.
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ library_name: transformers
4
+ base_model: PaddlePaddle/PP-OCRv6_medium_rec_safetensors
5
+ tags:
6
+ - captcha
7
+ - ocr
8
+ - minecraft
9
+ - anti-bot
10
+ - pp-ocr
11
+ - image-to-text
12
+ - pytorch
13
+ pipeline_tag: image-to-text
14
+ ---
15
+
16
+ # CaptchaOCR PP-OCRv6 β€” a text recogniser for Minecraft anti-bot captchas
17
+
18
+ > ## Status: the weights are **not published yet**. This repository holds the model card only.
19
+ >
20
+ > The fine-tune on the [CaptchaOCR dataset](https://huggingface.co/datasets/Titlehhhh/captchaocr-dataset)
21
+ > has **not been run to completion yet**, so there is **no checkpoint, no `config.json` and no
22
+ > measured metric** in this repository. Nothing here is invented and nothing is extrapolated from a
23
+ > probe run: the numbers quoted below are the **measured numbers of the base model before
24
+ > fine-tuning**, and they are labelled as such.
25
+ >
26
+ > **RU:** Ρ€Π΅ΠΏΠΎΠ·ΠΈΡ‚ΠΎΡ€ΠΈΠΉ ΠΌΠΎΠ΄Π΅Π»ΠΈ ΠΏΠΎΠΊΠ° **пуст ΠΏΠΎ вСсам** β€” здСсь Ρ‚ΠΎΠ»ΡŒΠΊΠΎ ΠΊΠ°Ρ€Ρ‚ΠΎΡ‡ΠΊΠ°. ΠžΠ±ΡƒΡ‡Π΅Π½ΠΈΠ΅ Π½Π°
27
+ > датасСтС `Titlehhhh/captchaocr-dataset` Π΅Ρ‰Ρ‘ Π½Π΅ Π·Π°Π²Π΅Ρ€ΡˆΠ΅Π½ΠΎ, поэтому **ΠΌΠ΅Ρ‚Ρ€ΠΈΠΊΠΈ появятся послС
28
+ > ΠΏΡ€ΠΎΠ³ΠΎΠ½Π°** (плСйслолдСр Π½ΠΈΠΆΠ΅ β€” это Π½Π΅ Π²Ρ‹Π΄ΡƒΠΌΠ°Π½Π½Ρ‹Π΅ числа). Π’Ρ‹ΠΊΠ»Π°Π΄ΠΊΠ° ΠΏΡ€ΠΎΠ±Π½ΠΎΠ³ΠΎ Ρ‡Π΅ΠΊΠΏΠΎΠΈΠ½Ρ‚Π° (2 эпохи)
29
+ > ΡΠΎΠ·Π½Π°Ρ‚Π΅Π»ΡŒΠ½ΠΎ Π½Π΅ сдСлана: ΠΎΠ½ Ρ…ΡƒΠΆΠ΅ Π±Π°Π·ΠΎΠ²ΠΎΠΉ ΠΌΠΎΠ΄Π΅Π»ΠΈ.
30
+
31
+ ## What it reads
32
+
33
+ Frames from the [CaptchaOCR dataset](https://huggingface.co/datasets/Titlehhhh/captchaocr-dataset) β€” the kind of
34
+ pictures the model is trained on. The caption gives the type and the **gold answer** from
35
+ `labels.jsonl` (what the player has to type). The links are absolute: the pictures live in the
36
+ dataset repository, not in this one.
37
+
38
+ | | | |
39
+ |---|---|---|
40
+ | ![Minecraft map captcha](https://huggingface.co/datasets/Titlehhhh/captchaocr-dataset/resolve/main/data/out/haq/mapcha/map/001.png) | ![Latin code captcha](https://huggingface.co/datasets/Titlehhhh/captchaocr-dataset/resolve/main/data/out/LagBug/XProtect/code/001.png) | ![Cyrillic captcha](https://huggingface.co/datasets/Titlehhhh/captchaocr-dataset/resolve/main/data/out_ru/LimboFilter/symbols/001.png) |
41
+ | `haq/mapcha/map` β€” 128Γ—128 map, 4 Latin characters Β· answer `s8TO` | `LagBug/XProtect/code` β€” 130Γ—130, white background, noise Β· answer `EzhYNY` | `LimboFilter/symbols` β€” 128Γ—128, 3 Cyrillic characters Β· answer `ΠͺМИ` |
42
+
43
+ ## What this model is
44
+
45
+ A fine-tune of **PP-OCRv6 medium** (PaddleOCR's text recogniser) on the CaptchaOCR dataset: the
46
+ captchas that open-source Minecraft anti-bot plugins draw, rendered by this project's own Java
47
+ generators (46 types, 4600 labelled PNG).
48
+
49
+ * **Architecture:** PP-OCRv6 medium recogniser β€” PP-LCNetV4 backbone + CTC head,
50
+ `transformers` class `PPOCRV6SmallRecForTextRecognition`
51
+ (`model_type: pp_ocrv6_small_rec`), 19.1 M parameters, hidden size 192.
52
+ * **Dictionary:** the base model's own 18 710 classes (index 0 = CTC blank, space at the end).
53
+ * **Input:** the captcha **text band** (the strip of pixels that carries the answer), not the whole
54
+ frame: resized to height 48, width scaled proportionally and right-padded to 320, BGR channel
55
+ order, normalised as `(x / 255 - 0.5) / 0.5`.
56
+ * **Output:** per-timestep class probabilities; decoded greedily (CTC: collapse repeats, drop the
57
+ blank). Greedy decoding is the right decoder here β€” the answers are random strings, so a language
58
+ model has nothing to exploit.
59
+
60
+ **The band is the whole game.** On the 29 dataset types that carry a band, the base model scores
61
+ `acc_micro = 0.5517` on the band and `acc_micro = 0.0345` on the whole frame β€” about 16Γ— worse on
62
+ the frame. A solver must crop first.
63
+
64
+ ## Provenance and licence
65
+
66
+ * **Base model:** [`PaddlePaddle/PP-OCRv6_medium_rec_safetensors`](https://huggingface.co/PaddlePaddle/PP-OCRv6_medium_rec_safetensors)
67
+ β€” Apache-2.0, by PaddlePaddle.
68
+ * **This model:** that base, fine-tuned on the CaptchaOCR dataset published here:
69
+ [`Titlehhhh/captchaocr-dataset`](https://huggingface.co/datasets/Titlehhhh/captchaocr-dataset)
70
+ (46 captcha types Β· 4600 labelled PNG Β· 23 upstream plugins).
71
+ * **Licence of the weights:** Apache-2.0, inherited from the base model. The fine-tuning code and
72
+ this card are released under the same terms.
73
+ * **Licence of the dataset:** *not* Apache-2.0 β€” the dataset is split by the licence of the
74
+ upstream plugin of each captcha type. The dataset card has the per-type table.
75
+
76
+ ## Measured results
77
+
78
+ ### This repository's weights
79
+
80
+ **No metrics yet β€” the training run has not been completed.** This section will hold exact match
81
+ and CER per type and on the type-held-out split when the fine-tune is done. Until then, treat the
82
+ numbers below as the **starting point**, not as the result of this repository.
83
+
84
+ ### The base model, measured before fine-tuning (baseline)
85
+
86
+ These are measured numbers of `PaddlePaddle/PP-OCRv6_medium_rec_safetensors` **as it is**, on the
87
+ CaptchaOCR dataset β€” i.e. the point the fine-tune starts from:
88
+
89
+ | measurement | value | source |
90
+ |---|---|---|
91
+ | exact match (micro) on the 7 type-held-out types, band crop | 0.5586 | `train/run_ppocr/` zero-step baseline |
92
+ | exact match (median over those 7 types) | 0.60 | `train/run_ppocr/` zero-step baseline |
93
+ | CER (micro) on the same split | 0.2561 | `train/run_ppocr/` zero-step baseline |
94
+ | exact match (micro) on the 29 band-carrying types, band crop | 0.5517 | `reports/bench_models/ppocrv6-medium.json` |
95
+ | exact match (micro) on the same 29 types, whole frame | 0.0345 | `reports/bench_models/ppocrv6-medium.json` |
96
+
97
+ Both sources are files in the **CapthaModel** project repository; nothing above is a projection.
98
+
99
+ **Why the weights are not here yet.** A short probe run (2 epochs) of the fine-tune was measured
100
+ on the same held-out types and it came out **worse than the base model** β€” median exact match
101
+ 0.41 after epoch 1 and 0.24 after epoch 3, against 0.60 for the untouched base model. That
102
+ checkpoint is deliberately **not** published: a two-epoch probe is not a model, and shipping it as
103
+ one would misrepresent it.
104
+
105
+ ## How to use it
106
+
107
+ > The snippet below is the **preprocessing the project itself uses** (`train/ppocr_data.py`), and it
108
+ > was verified end-to-end against the base model on a real sample of the dataset (it decoded the
109
+ > held-out sample `JSH32/Minecaptcha/map/001.png`, band `[25, 77, 79, 35]`, as `PTAY`, which is the
110
+ > gold answer). It is written for the moment the weights are published β€” until then
111
+ > `from_pretrained("Titlehhhh/captchaocr-ppocrv6")` has nothing to load.
112
+
113
+ ```python
114
+ import json
115
+
116
+ import numpy as np
117
+ import torch
118
+ from huggingface_hub import hf_hub_download
119
+ from PIL import Image
120
+ from transformers import PPOCRV6SmallRecForTextRecognition
121
+
122
+ REPO_ID = "Titlehhhh/captchaocr-ppocrv6"
123
+
124
+ model = PPOCRV6SmallRecForTextRecognition.from_pretrained(REPO_ID).eval()
125
+ # the dictionary comes with the model repository (preprocessor_config.json)
126
+ character_list = json.load(
127
+ open(hf_hub_download(REPO_ID, "preprocessor_config.json"), encoding="utf-8")
128
+ )["character_list"]
129
+
130
+ HEIGHT, MAX_WIDTH = 48, 320
131
+
132
+
133
+ def prepare(band: Image.Image) -> torch.Tensor:
134
+ """band -> (1, 3, 48, 320) float tensor."""
135
+ scale = HEIGHT / max(1, band.height)
136
+ w = max(1, min(MAX_WIDTH, int(round(band.width * scale))))
137
+ small = band.convert("RGB").resize((w, HEIGHT), Image.BILINEAR)
138
+ a = np.asarray(small, dtype=np.float32) / 255.0
139
+ a = a[:, :, ::-1] # BGR, as PP-OCRv6 expects
140
+ a = (a - 0.5) / 0.5 # paddle normalisation
141
+ canvas = np.zeros((3, HEIGHT, MAX_WIDTH), dtype=np.float32)
142
+ canvas[:, :, :w] = a.transpose(2, 0, 1)
143
+ return torch.from_numpy(canvas)[None]
144
+
145
+
146
+ def greedy_decode(probs) -> str:
147
+ """CTC greedy decode: collapse repeats, drop the blank."""
148
+ best = probs.argmax(dim=-1)[0].tolist()
149
+ chars, prev = [], -1
150
+ for k in best:
151
+ if k != prev and k != 0:
152
+ chars.append(character_list[k])
153
+ prev = k
154
+ return "".join(chars)
155
+
156
+
157
+ image = Image.open("captcha.png")
158
+ x, y, w, h = (25, 77, 79, 35) # the answer band: take it from labels.jsonl
159
+ band = image.crop((x, y, x + w, y + h))
160
+
161
+ with torch.no_grad():
162
+ probs = model(pixel_values=prepare(band)).last_hidden_state
163
+
164
+ print(greedy_decode(probs))
165
+ ```
166
+
167
+ The band (`roi`) is not optional: on scene-like captchas the answer is one line among several, and
168
+ the whole frame makes the model read the wrong line.
169
+
170
+ ## Reproducing
171
+
172
+ The training code lives in the **CapthaModel** repository:
173
+
174
+ ```powershell
175
+ .venv-gpu\Scripts\python.exe train\finetune_ppocr.py --epochs 30 --batch-size 32
176
+ ```
177
+
178
+ `train/finetune_ppocr.py` fine-tunes `PaddlePaddle/PP-OCRv6_medium_rec_safetensors` on the
179
+ dataset, holds out whole **plugins** (never random splits β€” a random split leaks the exact gold
180
+ answer of 34 % of the test samples), and reports exact match and CER per type plus the median over
181
+ types. `train/ppocr_data.py` holds the dataset loader and the preprocessing shown above.
182
+
183
+ ## Limitations
184
+
185
+ * **Only the text band is read.** This is a recogniser, not a solver: it does not reverse strings,
186
+ does not turn Russian number words into digits and does not compute arithmetic.
187
+ * **No Cyrillic in the dictionary.** The base model's 18 710 classes contain no Cyrillic at all, so
188
+ the Cyrillic captcha types of the dataset cannot be expressed in this model's alphabet.
189
+ * **The data is synthetic.** Every image is a render of this project's own generators β€” faithful to
190
+ the upstream algorithms, but not a screenshot of a live server.
191
+ * **Generalisation is the hard part.** A recogniser can memorise a plugin's style instead of
192
+ learning the characters; only the type-held-out numbers mean anything.
193
+
194
+ ## Citation
195
+
196
+ This model is a fine-tune of PP-OCRv6 by PaddlePaddle (Apache-2.0). The dataset is the
197
+ CaptchaOCR dataset, re-rendered from open-source Minecraft anti-bot plugins. See the dataset card
198
+ for the provenance and the per-type licence of every captcha type.