output_dir = /data/horst_checkpoints (direct bucket writes)

Add /data bucket sync on startup for HF Space

Checkpoints: steps-based save, best model loading, auto-sync to bucket on each save

Add bucket checkpoint sync after training

Update README.md

Fix Gradio 6.0: move theme to launch(), add missing launch()

Light UI theme: clean two-panel layout, base theme

Simplify Gradio UI: left/right split layout, minimal design

Update: 10000 Trainingsdaten, HF_TOKEN Secret, README fix

Restore full app.py with training and model loading

UI preview: lightweight Gradio serving custom HTML

Redesign: LanguageTool-inspired single-column UI with checkboxes, word diff, and changes list

Fix: name error (do_training order), gradio 6.0 deprecation, 401 hub check

Smart auto-training: check for existing model, only train if needed, guard against restart

Professional UI: Training im Accordion, klare Hierarchie, Beispieltext, Korrekturmodi funktional

Move css/theme to launch() for Gradio 6

Fix Trainer API: processing_class, warmup_steps

Clean modern UI: links Eingabe, rechts Ausgabe

Remove predict_with_generate arg

Fix Gradio 6 API: remove show_copy_button, move theme/css to launch

Premium UI + echte Korrektur-Logik + Training

Remove overwrite_output_dir arg

Fix Gradio API, add page refresh

Add on-device CPU training pipeline

Use T5Tokenizer + T5ForConditionalGeneration directly

Replace with pipeline-based app using oliverguhr/spelling-correction-german-base

Add sentencepiece to requirements.txt

Use slow tokenizer (no sentencepiece needed)

Fix: use gbert-base directly, add sentencepiece dep

Fix short_description length

Add Rechtschreibkorrektur app with gbert-base MLM

initial commit