metadata

title: RestrictedPython Code Eval
datasets:
  - N/A (eval module only)
tags:
  - evaluate
  - metric
description: >-
  Same logic as the built-in `code_eval`, but compiling and running the code
  using `RestrictedPython`
sdk: gradio
sdk_version: 3.19.1
app_file: app.py
pinned: false

Metric Card for RestrictedPython Code Eval

Metric Description

A code-based evaluation metric, with the same logic as code_eval.

How to Use

from evaluate import load
code_eval = load("guydav/restrictedpython_code_eval")
test_cases = ["assert add(2,3)==5"]
candidates = [["def add(a,b): return a*b", "def add(a, b): return a+b"]]
pass_at_k, results = code_eval.compute(references=test_cases, predictions=candidates, k=[1, 2], use_safe_builtins=True)

N.B. This metric exists to run untrusted model-generated code. Users are strongly encouraged not to do so outside of a robust security sandbox. Before running this metric and once you've taken the necessary precautions, you will need to set the HF_ALLOW_CODE_EVAL environment variable. Use it at your own risk:

import os
os.environ["HF_ALLOW_CODE_EVAL"] = "1"`

Inputs

The following arguments are inherited from the basic code_eval:

predictions (List[List[str]]): a list of candidates to evaluate. Each candidate should be a list of strings with several code candidates to solve the problem.

references (List[str]): a list with a test for each prediction. Each test should evaluate the correctness of a code candidate.

k (List[int]): number of code candidates to consider in the evaluation. The default value is [1, 10, 100].

num_workers (int): the number of workers used to evaluate the candidate programs (The default value is 4).

timeout (float): The maximum time taken to produce a prediction before it is considered a "timeout". The default value is 3.0 (i.e. 3 seconds).

In addition, this metric supports three additional arguments, specifying which imports should be made available, and controlling other apsects of RestrictedPython behavior:

use_safe_builtins (bool): Whether or not to allow the usage of RestrictedPython.safe_builtins, defaults to True

use_limited_builtins (bool): Whether or not to allow the usage of RestrictedPython.limited_builtins, which provides limited implementations of range, list, and tuple. defaults to True.

use_utility_builtins (bool): Whether or not to allow the usage of RestrictedPython.utility_builtins, which includes the string, math, random, and set packages, among others. Defaults to True.

additional_globals (Dict[str, Any] | None): Any additional globals to make available to the code. Defaults to None.

additional_locals (Dict[str, Any] | None): Any additional locals to make available to the code. Defaults to None.

allowed_imports (List[str] | None): A list of allowed imports. Defaults to None.

As the new arguments are optional, this could be used as a drop-in replacement for code_eval.

Additionally, this metric sets several different globals if they are not provided as additional globals. The full list of globals set is: __metaclass__, __name__, _getiter_, _iter_unpack_sequence_, _getitem_, getattr, _write_, _inplacevar_, _print_. See the code for additional details.

Output Values

Identical to code_eval:

The Code Eval metric outputs two things:

pass_at_k: a dictionary with the pass rates for each k value defined in the arguments.

results: a dictionary with granular results of each unit test.

Values from popular papers

The original CODEX paper reported that the CODEX-12B model had a pass@k score of 28.8% at k=1, 46.8% at k=10 and 72.3% at k=100. However, since the CODEX model is not open source, it is hard to verify these numbers.

Examples

Copied from the code_eval model card:

Full match at k=1:

from evaluate import load
code_eval = load("code_eval")
test_cases = ["assert add(2,3)==5"]
candidates = [["def add(a, b): return a+b"]]
pass_at_k, results = code_eval.compute(references=test_cases, predictions=candidates, k=[1])
print(pass_at_k)
{'pass@1': 1.0}

No match for k = 1:

from evaluate import load
code_eval = load("code_eval")
test_cases = ["assert add(2,3)==5"]
candidates = [["def add(a,b): return a*b"]]
pass_at_k, results = code_eval.compute(references=test_cases, predictions=candidates, k=[1])
print(pass_at_k)
{'pass@1': 0.0}

Partial match at k=1, full match at k=2:

from evaluate import load
code_eval = load("code_eval")
test_cases = ["assert add(2,3)==5"]
candidates = [["def add(a, b): return a+b", "def add(a,b): return a*b"]]
pass_at_k, results = code_eval.compute(references=test_cases, predictions=candidates, k=[1, 2])
print(pass_at_k)
{'pass@1': 0.5, 'pass@2': 1.0}

Limitations and Bias

From the original code_eval model card:

As per the warning included in the metric code itself:

This program exists to execute untrusted model-generated code. Although it is highly unlikely that model-generated code will do something overtly malicious in response to this test suite, model-generated code may act destructively due to a lack of model capability or alignment. Users are strongly encouraged to sandbox this evaluation suite so that it does not perform destructive actions on their host or network. For more information on how OpenAI sandboxes its code, see the accompanying paper. Once you have read this disclaimer and taken appropriate precautions, uncomment the following line and proceed at your own risk:

More information about the limitations of the code can be found on the Human Eval Github repository.

Additionally, this metric does not currently allow for custom RestrictedPython policies -- so any code that depends on non-default libraries or packages may fail for that reason.

TODO: Add a use_custom_builtins argument that allows users to specify their own RestrictedPython policy. See the RestrictedPython documentation for additional details.

Citation

Based on the original code_eval metric, which cites:

@misc{chen2021evaluating,
      title={Evaluating Large Language Models Trained on Code},
      author={Mark Chen and Jerry Tworek and Heewoo Jun and Qiming Yuan \
and Henrique Ponde de Oliveira Pinto and Jared Kaplan and Harri Edwards \
and Yuri Burda and Nicholas Joseph and Greg Brockman and Alex Ray \
and Raul Puri and Gretchen Krueger and Michael Petrov and Heidy Khlaaf \
and Girish Sastry and Pamela Mishkin and Brooke Chan and Scott Gray \
and Nick Ryder and Mikhail Pavlov and Alethea Power and Lukasz Kaiser \
and Mohammad Bavarian and Clemens Winter and Philippe Tillet \
and Felipe Petroski Such and Dave Cummings and Matthias Plappert \
and Fotios Chantzis and Elizabeth Barnes and Ariel Herbert-Voss \
and William Hebgen Guss and Alex Nichol and Alex Paino and Nikolas Tezak \
and Jie Tang and Igor Babuschkin and Suchir Balaji and Shantanu Jain \
and William Saunders and Christopher Hesse and Andrew N. Carr \
and Jan Leike and Josh Achiam and Vedant Misra and Evan Morikawa \
and Alec Radford and Matthew Knight and Miles Brundage and Mira Murati \
and Katie Mayer and Peter Welinder and Bob McGrew and Dario Amodei \
and Sam McCandlish and Ilya Sutskever and Wojciech Zaremba},
      year={2021},
      eprint={2107.03374},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}

Further References

The original code_eval metric: https://huggingface.co/spaces/evaluate-metric/code_eval
RestrictedPython: https://restrictedpython.readthedocs.io/en/latest/index.html