Instruction Following Eval

This is an unofficial fork of IFEval (see paper), designed to make running instruction following eval easier. The goals of the fork are as follows:

• Make the codebase pip installable & fix the dependencies
• Provide runtime bindings that can be called from your own codebase. If you are interested in a CLI based interface see the original project.

Install

You can install instruction_following_eval with pip directly

pip install git+https://github.com/josejg/instruction_following_eval.git

or manually cloning it:

git clone https://github.com/josejg/instruction_following_eval.git
cd instruction_following_eval
pip install . 

Running IFEval

To run instruction following eval in your codebase you just need to import instruction_following_eval and provide a list of examples encoded as dictionaries with the following structure:

# Example test
{
 'key': 1001,
 'instruction_id_list': ['punctuation:no_comma'],
 'prompt': 'I am planning a trip to Japan, and I would like thee to write an '
           'itinerary for my journey in a Shakespearean style. You are not '
           'allowed to use any commas in your response.',
 'kwargs': [{}],
 'response': '<MODEL RESPONSE>'
}

However, you can retrieve the list from examples of the original paper by calling default_examples.

from instruction_following_eval import default_examples, instruction_following_eval

examples = default_examples()

for example in examples:
    example['response'] = model.generate(example['prompt'])

accuracy = instruction_following_eval(examples)

print(accuracy)

Tests

$ python test/instructions_test.py
$ python test/instructions_util_test.py