--- language: - de license: apache-2.0 task_categories: - token-classification pretty_name: German ticket privacy data configs: - config_name: default data_files: - split: train path: data/train.jsonl - split: validation path: data/validation.jsonl - split: test path: data/test.jsonl --- # German ticket privacy data Fictional German support questions, replies and human-specified privacy spans for supervised token classification. The train split has 4,200 rows, development has 420 and the independently authored public test has 121. The training corpus covers 15 privacy labels. The field `context` is the preceding support question, `text` is the reply and `gold` contains half-open Unicode code-point offsets into that reply. Empty context denotes standalone text. Pairs deliberately reuse the same reply under a different question. Private-looking words can refer to public titles or catalogue items. All records are synthetic. The train and development generators use separate value and wording banks. The independent test contains 28 matched pairs, 56 clean replies and 73 sensitive mentions. It was sealed before scoring this version. Once a model is tuned using this public test, evaluate it on a new independent panel. [Evaluated checkpoint and full scores](https://huggingface.co/EasyOctopus/german-ticket-privacy-filter)