suffix.GCGCached always runs all stored cached prompts
run.soft_probe_prompt_cap sets a cap on how many distinct prompts to offer per probe
suffix.GCGCached should not issue more prompts than run.soft_probe_prompt_cap
To see the issue, try a config like:
{
"plugins": {
"probe_spec": "suffix"
},
"run": {
"soft_probe_prompt_cap": 10,
"generations": 1,
}
}
Note that >10 prompts are issued
When addressing this, use a mechanism for randomly selecting prompts like that in probes.base._prune_data()