Implement MMLU_Evaluator.run()

[alinaryan]

This PR updates the MMLUEvaluator child class to include the lm-evaluation-harness lib dependency and adds an API call to simple_evaluate() to run MMLU.

ToDo:

• test running in a VM
• observe the output of results to help decide how to access individual scores vs. overall score
• finish writing the helper function to calculate the overall score based on the results
• parameterize these 57 tasks (I think CLI needs to do this?) https://github.com/instructlab/evaluation/blob/main/scripts/run_mmlu.sh#L24

[alimaredia]

If anyone needs to test this PR locally, after initializing the virtual env in the repo and installing all the dependencies this python script executes the function:

from instructlab.eval.mmlu import MMLU_Evaluator

tasks = ["mmlu_anatomy","mmlu_astronomy"]

# this path is to the granite-7b-lab model on huggingface not on the local filesystem 
model_path = "instructlab/granite-7b-lab"
mmlu = MMLU_Evaluator(model_path, tasks)
overall_score, individual_scores = mmlu.run()
print(overall_score)
print(individual_scores)

[JamesKunstle]

very nice PR, makes sense to me. couple of nits.

[alimaredia]

I've mentioned this in some of my comments in code review, Getting mmlu working would be a big deal for the library so I propose we do the following and then merge the PR ASAP:

1. Remove the testing bits from mmlu.py
2. Someone besides the authors (Alina or I) test the PR by hand, to see the scores being output properly
3. Create follow up issues for anything else that we know needs to be done.

[cdoern]

missing model path, run tox -e ruff as well! looks good though, I will approve after those changes

[nathan-weinberg]

@alinaryan can you rebase?

[danmcp]

Nit: Use f""

[cdoern]

LGTM