Order prompts within probes for max bag attack success rate

[leondz]

Summary

For probes that don't compose their prompts, we can get info on which prompts are more or less successful in the bag of models used for calibration. This can be used to order prompts within a probe.

Once we have this infrastructure, we can then apply it to probes that compose their prompts out of components, as long as the component choice is logged in the attempt (not in this feature).

Process

• update the calibration process to record (probe, detector, prompt, asr) tuples
• from this derive (probe, prompt, asr) tuples taking the mean (not median) over all detectors used
• work out a way of storing this artefact (HF? compressed and local? use a small hash like crc32? cut out everything below a certain ASR?)
• work out where to add this function (probes.base.Probe() ?)
• implement the sort
• add a config variable for whether to do this optimisation