Overview

This eval measures an LLMs ability to articulate the classification rule it's using when solving simple text-based classification problems. Specifically, it contains two categories of tasks:

Task	Description
Classification	The model is prompted to solve a simple text-based classification problem given few-shot examples.
Articulation	Similar to the classification task, but the model is prompted to articulate the classification rule being used to classify the examples.

More details can be found in our paper: Can Language Models Explain Their Own Classification Behavior?

Setup

To run our eval, please clone the OpenAI evals repo and follow the installation instructions.

Within the OpenAI evals repo, you can run the in-distribution variant of Articulate Rules with:

oaieval <model> articulate-rules.<task>.in-distribution.<rule>

where:

• task is either one of classification or articulation
• rule is a valid rule (e.g. contains_a_digit)

Similarly, the out-of-distribution variants can be run with:

oaieval <model> articulate-rules.<task>.out-of-distribution.<rule>.<attack>

where:

• task is either one of classification or articulation
• rule is a valid rule (e.g. contains_a_digit)
• attack is a valid adversarial attack (e.g. SynonymAttackBehaviour)

The list of valid task, rule and attack combinations are defined in the Articulate Rules yaml file.

Customiziation

We've provided two scripts, scripts/create-eval.py and scripts/create-eval-spec.py, which allow you to create custom variants of our eval in the format of an OpenAI eval. This may be useful if you wish to increase the number of few-shot examples, for example. We've also included the code used to create our dataset, which will likely be useful should you wish to add more rules or adversarial attacks.