Add annotator test runner + LlamaGuard2, Llama 3 70b annotator test

[tsunamit]

No description provided.

[github-actions]

MLCommons CLA bot All contributors have signed the MLCommons CLA ✍️ ✅

[bkorycki]

This is a good start! The main issue I see is in how the test creates test items and how LlamaGuard2SUT processes them.

• Right now the test produces ChatPrompts, but LlamaGuard2SUT can only handle TextPrompts.
• There are two layers of prompt formatting (in the test and by LlamaGuardAnnotator's default formatter in translate_request())

I think we can simply things by pulling all of the formatting into LlamaGuard2SUT. This would involve:

• The test produces TextPrompt(text=assistant_response)
• Define a custom formatting function (if you don’t want to use LlamaGuardAnnotator’s default formatter).
• Initialize LlamaGuard2SUT’s llama_guard_client with your custom formatter.

[tsunamit]

@bkorycki question regarding chat vs text prompts.

1. llama guard 2 as a safety model produces text responses
2. we have other safety models (llama 3 70b, mistral) that product chat responses

Is it possible to have a single test handle both? or do I need to create 2 different tests

The only difference I recall making was to swap requrest_sut_capabilities

(edit) Oh shoot, realized there are several locations where I had to change it to Chat objects

[bkorycki]

Right now tests have to choose between producing test items that either have TextPrompts or ChatPrompts. It is a SUT's responsibility to handle the different prompt types by implements translate_text_prompt and/or translate_chat_prompt. If a model only accepts chat prompts, then you can you just transform the test's TextPrompt into chat format in translate_text_prompt.

[tsunamit]

Here's the refactor and proposed design decision so far

Decisions and assumptions

1. Safety Models are SUTs- decided to keep this paradigm for now. Creating a new runner and new AnnotationTestItem base classes is a viable approach, but one I'm electing to defer for simplicity
2. Create 2 separate test types - LlamaGuard tests and Chat model tests are just 2 separate tests. They share reusable logic via util functions

Deferred + to address in future

1. Disentangling lg2 annotator and lg2 SUT, creating a more logical relationship if sharing code
2. Refactor entire system to use new runner for annotator evals

[bkorycki]

I guess my concern with this approach is that tests are built/applied to only one type of a safety SUT. So for every new evaluator we want to test, we need to create 1) a new SafetyModelSUT to transform the annotator into a SUT and 2) a new test with custom processing of the model’s response.

This process increases the amount of effort required to actually use the framework. So --and maybe I'm missing something-- I’m not sure what benefits are gained by awkwardly forcing evaluators in SUT classes.

Using a custom runner would allow new evaluators to be tested without any additional work. Setting this up would just require one new test and the custom runner, which only needs a few modifications to the simple_test_runner.

[tsunamit]

Good points. After refactoring, here are the design decisions

Design decisions

1. Safety models are Annotators.
2. Evaluation of annotators uses it's own annotator runner. No longer need to create SUTs that map to Annotator. Thanks to runner separation, we don't need to worry about conflicting with SUT tests, they're completely separate

Tests and annotators supported

1. Safety model test for LlamaGuard2 Annotator
2. Safety model test for chat models e.g. Llama 3 70b chat

Todo

1. Add Mistral 8x22b instruct annotator

[bkorycki]

I think this looks a lot nicer! I think there's just one high-level issue remaining before we can merge this in: being able to run one test class on multiple annotators. This test always expects a specific interface to an annotation,SafetyModelResponse, regardless of the annotator. Therefore it makes more sense to make the annotators responsible for translating their native annotation into the expected SafetyModelResponse.
This involves:

• Annotators to "test" are specified at runtime. The runner accepts an annotator(s) param, parallel to the simple runner's SUT param. The test does not return anything in get_annotators().
• Annotators must return a SafetyModelResponse. We can add a formal check in the runner in the future.
  • In order to test the LlamaGuardAnnotator, we need to just add a new class that extends it and only overrides it's translate_response(). Should be just a couple lines.
• Only use one test class.

[bkorycki]

I don't think this test requires any api keys now.