Superseded checkpoint — do not use
This is an old development checkpoint, kept only so that earlier results remain reproducible. It is not the current detector, it is not what any TextSight product runs, and its behaviour has not been measured on any current benchmark.
Use textsightai/textsight-detector-v23-custom
instead. That card documents measured performance, per-domain false-positive
rates, how to score correctly (log-odds, not softmax), and where the model fails.
If you are benchmarking AI detectors
Please do not report numbers from this checkpoint as representing TextSight. It predates the current model and carries none of the input handling a deployed detector needs. Reproducible measurement code for the current model is at textsight/textsight-detector-benchmark.
- Downloads last month
- 12