Ethel Tutor Eval

This repository contains a framework for evaluating Ethel on education-specific benchmark datasets.

Datasets

Currently, our evaluation pipeline supports the following datasets:

• GSM8K
• MATH
• MGSM
• TutorEval

Scripts

For using the GSM8K, MATH or MGSM dataset with the evaluation pipeline

Use the following command with the dataset and model command line arguments to run the evaluation

Required arguments:

• --dataset: The dataset to be used for the evaluation, e.g.: GSM8K
• --model: The model type to generate the answer, e.g.: Ethel

Optional arguments:

• -- model_name: The exact name of the model, e.g.: swissai/ethel-70b-magpie
• -- limit: If you prefer to run the evaluation on a subset of the dataset
• -- n_shot: If you want to perform n-shot learning

Example:

python3 -m scripts.run_pipeline --dataset=MATH --model=Ethel

For running the TutorEval dataset with the evaluation pipeline

Required arguments:

• -- dataset: The dataset to be used for the evaluation, e.g.: GSM8K
• -- model: The model type that works as a tutor to generate the answer, e.g.: Ethel
• -- grader_model: The model type to grade the tutor's answer, e.g.: Ethel

Optional arguments:

• -- model_name: The exact name of the model, e.g.: swissai/ethel-70b-magpie
• -- grader_model_name: The exact name of the model, e.g.: swissai/ethel-70b-magpie
• -- limit: If you prefer to run the evaluation on a subset of the dataset
• -- closed_book: If you want to evaluate on a closed_book setup (or on open_book)

python3 -m scripts.run_tutoreval --dataset=TutorEval --model=Ethel --grader_model=Ethel

Authors:

This work was done as part of the ML4Science project at the CS-433 course at EPFL.

The main contributors are:

• Kamel Charaf
• Ivan Pavlov
• Michele Smaldone