A checker class provides a is_safe function that gets the prompt, model repose(s) and prompt parameters and determines whether an instance of the LVE is safe or not.
This is the main entry point where LVE runs checks to determines if model responses should be treated as safe or not for that instance. As often similar checks are required, LVE provides a set of Core Checkers.
This page further describes the types of checkers and how to specify custom checkers.
Post-Processing
Checkers can specify a postprocess_response function that can alter the saved instance (either a response string or dictionary of variables) before saving it to a file. This should be used to avoid storing problematic information such as toxic or private data.
For an example see the personally identifying information (PII) checker.
Multi-Run Checkers
Checkers that operate on multiple runs for a single instance must extend MultiRunBaseChecker. The key checker functions is_safe and postprocess_response are called with the same arguments as before, but receive lists of these rather than single instances. Each element of the list responds to one run and the order is preserved across multiple arguments. For an example see the MultriRunLambdaChecker.
Specifying Custom Checkers
We recommend to use or adapt the existing Core Checkers. where possible. However, if a custom checkers is required, there are two possibilities:
Add the checker to the core of lve-tools. Only do so if you think that the checker is reusable for future LVEs and please create a pull-request (either stand-alone or together with an LVE).
Package your custom checker with your LVE. This is the recommended approach if the checker is specific to your LVE. A single-file checker can be defined inside checker/checker.py inside an LVE. For an example see the LVE responsibility/copyright/songs/translation.
Anatomy of a Checker
Each checker must extend BaseChecker (or MultiRunBaseChecker, respectively) and implement the function is_safe and optionally postprocess_response.
The is_safe(self, prompt_out, param_values) function receives the prompt, including the model responses, and prompt parameters is available.
To extract the the model response from the prompt we provide the helper function extract_response_from_prompt.
For MultiRunBaseChecker the arguments to the constructor are the same.
The corresponding signature of is_safe is also similar, but prompt_out is a list of model responses, rather than a single model response.
