v1.0.0 — iKaptcha initial release

iKaptcha v1.0.0

First tagged release of the iKaptcha CRNN captcha solver.

Model

CRNN (CNN + BiLSTM + CTC), 1.66M parameters, trained on ~65k synthetic
and ~11k real Ikariam pirate-fortress captchas. Beats the YOLOv8n
baseline by 16 points on the original validation set.

Model	Params	Original val	Corrected val
YOLOv8n (IkabotAPI baseline)	~3 M	78.7%	81.2%
CRNN (this release)	1.66 M	95.0%	97.3%

Character accuracy: 99.5%.

Assets

Model files (standalone downloads — also bundled in the repo):

• ikaptcha.onnx — production ONNX model, 6.4 MB, opset 18. Loadable
  via onnxruntime, cv2.dnn (≥ 4.8.0), onnxruntime-web, etc.
• ikaptcha.pth — same weights in PyTorch format
• yolov8n-ikariam-pirates-mAP-0_989.onnx — the original YOLO baseline
  from IkabotAPI, kept here for apples-to-apples comparison

Training datasets (available as .tar.gz for Linux/macOS and .zip
for Windows):

• ikariam_pirate_captcha_dataset.{tar.gz,zip} — original 1,200/300
  YOLO dataset from IkabotAPI
• dataset_pseudo_v2.{tar.gz,zip} — production 11,210/298
  train/corrected-val set (original YOLO data + 99 hand-labeled +
  ~9.9k pseudo-labeled at ≥0.99 confidence)

Unix users can fetch both datasets with one command from the repo root:
bash scripts/download_data.sh