Ecosyste.ms: Awesome
An open API service indexing awesome lists of open source software.
https://github.com/danesherbs/articulate-rules
Can large language models explain their own classification behavior?
https://github.com/danesherbs/articulate-rules
Last synced: 3 days ago
JSON representation
Can large language models explain their own classification behavior?
- Host: GitHub
- URL: https://github.com/danesherbs/articulate-rules
- Owner: danesherbs
- License: mit
- Created: 2024-03-30T01:01:21.000Z (9 months ago)
- Default Branch: main
- Last Pushed: 2024-03-31T04:16:36.000Z (9 months ago)
- Last Synced: 2024-11-09T16:47:48.513Z (about 1 month ago)
- Language: Python
- Size: 63.5 KB
- Stars: 1
- Watchers: 2
- Forks: 1
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
- Citation: CITATION.cff
Awesome Lists containing this project
README
# 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?](www.broken-link.com)
# Setup
To run our eval, please clone the OpenAI evals [repo](https://github.com/openai/evals) and follow the installation instructions.
Within the OpenAI evals repo, you can run the in-distribution variant of Articulate Rules with:
```bash
oaieval articulate-rules..in-distribution.
```
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:
```bash
oaieval articulate-rules..out-of-distribution..
```
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](https://github.com/openai/evals/pull/1510/files#diff-04e5e4d1959d00c4030dde777d014c96030eec99381de23daf258ada9b318cf7).
# 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.