why this is threshold above 0.9 in promptinjection
#1
by
LinkedIn
- opened
this is simple input
Good catch. We also noticed this issue with some simple prompts. We will change the model to JasperLS/deberta-v3-base-injection because it will produce below threshold response. Alternatively, we are considering this model too https://huggingface.co/benediktpri/finetuned_bert_injection?
There is a new model deployed, which doesn't produce that many false-positives
asofter
changed discussion status to
closed
Cool, Is the new one this https://huggingface.co/benediktpri/finetuned_bert_injection