Detector/Evaluator output indexing mismatch in multi-turn conversations

[saichandrapandraju]

There's a mismatch between what detectors return and what evaluators expect when dealing with multi-turn conversations:

• Detectors are expected to return a list of length len(all_outputs) (all assistant turns in the conversation)
• Evaluators index into attempt.outputs (only the last assistant turn's output)

The issue manifests in garak/evaluators/base.py:81 - where messages.append(attempt.outputs[idx]) assumes alignment with detector results. However detector results are of length attempt.all_outputs which is greater than attempt.outputs in multi-turn setting.

I wanted to check if this is an expected behavior..?

[leondz]

Thanks. This is an issue, you're right.

The current overall goal is to provide for more than one pattern for extracting target responses for detection (e.g. both last outputs, and all outputs, and other options). The pattern is currently expressed insufficiently clearly and we have an architecture todo on upgrading this. The intended situation is:

• Detectors are expected to return a list of length equal to that of .detect()'s input
• Detectors should also supply a list of the evidence used for each of detect()'s results
• Evaluators should not index attempts but instead detector evidence outputs

#1415 is thematically similar to this issue, and #1414 is pushing requirements in this area as well. We intend to tidy this all up before long. If there is a concrete bug then please say, that should be fixed before the more general architectural upgrade.

[saichandrapandraju]

Thank you, @leondz!
As long as the conversation in an Attempt has 1 assistant turn, there won't be any issue. But if it has more assistant turns, then the Evaluator will cause an index out of range error (when detector assigns hit for >0 index). #1415 is the solution for this.

Detectors are expected to return a list of length equal to that of .detect()'s input

May I know what you mean by "length equal to that of .detect()'s input", since .detect() accepts single Attempt as input? you mean length of 1?

[leondz]

Yes! An attempt can contain many outputs, and there have been various patterns for extracting these. For example, if _config.run.generations = 5, then for each prompt created by a probe, inference will be run five times, and each attempt will contain 5 outputs. So ".detect()'s input" refers to the number of inference results in the passed Attempt, sorry for the ambiguity.

[saichandrapandraju]

Got it, thank you!

[saichandrapandraju]

closing as #1415 fixed this