Notebooks

Executable notebooks that demonstrate the full txtcaptcha pipeline. The rendered versions below use cached outputs — to re-run them you need a GPU machine and the labeled datasets on disk (see download_dataset).

• Train unified model — download every labeled dataset, merge them into one folder and fit a single CRNN on the full alphanumeric vocabulary.
• Evaluate per dataset — per-source accuracy on a held-out split of the training corpus.
• Evaluate per dataset (live) — smoke-test the model on freshly downloaded, unlabeled captchas to check for overfit.