Mehran-NixiAI/treatment-status-de-authored

Treatment Status DE Authored

This dataset contains the authored arm of the German treatment-status benchmark. It is designed as a controlled minimal-pair dataset for deciding whether a mentioned treatment is currently given or not given.

Provenance

• Authored in-house for the benchmark study.
• Frozen as a public release for the authored arm only.
• The GraSCCo arm is excluded from this repo because it has separate licensing and provenance.

Files

• authored_eval_set.csv - frozen authored evaluation items
• id2family.csv - family mapping for bootstrap resampling and minimal-pair grouping
• annotation_guideline_v1.2.md - frozen label guideline
• preregistration_v1.md - study preregistration

Column dictionary for authored_eval_set.csv

Requested field names map to the actual CSV as follows:

• sentence -> text
• leaf -> leaf
• binary -> binary
• label_raw -> not a separate column in this release; the frozen label is already expressed in leaf
• phenomenon -> phen
• family -> join from id2family.csv
• source -> source

Actual CSV columns:

• id: stable item identifier
• source: item provenance label (authored)
• doc: source document identifier if present
• heading: local section heading context
• text: the clinical sentence or mention to classify
• leaf: frozen leaf label, such as GIVEN, PLANNED, DISCUSSED, REFUSED, WITHHELD, PAUSED, STOPPED, or HISTORICAL
• binary: collapsed label, given or not_given
• phen: phenomenon tag(s) used for analysis and oversampling

Notes:

• family is available through id2family.csv and is used for family-level bootstrap resampling.

Design

The authored set was built as a controlled minimal-pair benchmark so that a model must read the status language instead of relying on the drug name alone. The public white paper reports the exact family count, kappa, and safety metrics.

Labels

• given means the treatment is currently administered or actively continued.
• not_given means planned, discussed, refused, withheld, paused, stopped, or historical.

Companion repository

GitHub code and reproducibility pipeline:

• https://github.com/MehranNixiAi/clinical-treatment-status-benchmark