ssoad / human-eval-infilling

Code for the paper "Efficient Training of Language Models to Fill in the Middle"

HumanEval Infilling Benchmarks

This is an evaluation harness for the HumanEval infilling benchmarks described in the FIM paper.

Installation

Make sure to use python 3.7 or later:

$ conda create -n codex python=3.7
$ conda activate codex

Check out and install this repository:

$ git clone https://github.com/openai/human-eval-infilling
$ pip install -e human-eval-infilling

Usage

This program exists to run untrusted model-generated code. Users are strongly encouraged not to do so outside of a robust security sandbox. The execution call in execution.py is deliberately commented out to ensure users read this disclaimer before running code in a potentially unsafe manner. See the comment in execution.py for more information and instructions.

After following the above instructions to enable execution, generate samples and save them in the following JSON Lines (jsonl) format, where each sample is formatted into a single line like so:

{"task_id": "Corresponding task ID from the desired benchmark", "completion": "Completion only without the prompt"}

Ensure that the task_id used matches the task_id from the desired benchmark. See below and the paper for information on the benchmarks available.

We provide example_problem.jsonl and example_solutions.jsonl under data to illustrate the format and help with debugging.

Here is nearly functional example code (you just have to provide generate_one_completion to make it work) that saves generated completions for the single-line infilling benchmark to samples.jsonl.

from human_eval_infilling.data import write_jsonl, read_problems

problems = read_problems(benchmark_name="single-line")

num_samples_per_task = 100
samples = [
    dict(task_id=task_id, completion=generate_one_completion(problems[task_id]["prompt"], problems[task_id]["suffix"]))
    for task_id in problems
    for _ in range(num_samples_per_task)
]
write_jsonl("samples.jsonl", samples)

To evaluate the samples, run

$ evaluate_infilling_functional_correctness samples.jsonl --benchmark_name=single-line
Reading samples...
103300it [00:01, 23787.50it/s]
Running test suites...
100%|...| 103300/103300 [16:11<00:00, 33.76it/s]
Writing results to samples.jsonl_results.jsonl...
100%|...| 103300/103300 [00:00<00:00, 42876.84it/s]
{'pass@1': ..., 'pass@10': ..., 'pass@100': ...}

This script provides more fine-grained information in a new file ending in <input_path>_results.jsonl. Each row now contains whether the completion passed along with the execution result which is one of "passed", "timed out", or "failed".

As a quick sanity-check, the example samples should yield 30% pass@1.

$ evaluate_infilling_functional_correctness data/example_samples.jsonl --benchmark_name=test
Reading samples...
10it [00:00, 3365.94it/s]
100%|...| 10/10 [00:03<00:00,  2.76it/s]
Writing results to data/example_samples.jsonl_results.jsonl...
100%|...| 10/10 [00:00<00:00, 1309.08it/s]
{'pass@1': 0.30000000000000004, 'pass@10': 1.0}

There are 4 available benchmarks: single-line, multi-line, random-span, random-span-light. The first two are introduced in the InCoder paper and the latter two are introduced in the FIM paper. All benchmarks are used extensively in the FIM paper. There is also a dummy benchmark for testing.

Because there is no unbiased way of estimating pass@k when there are fewer samples than k, the script does not evaluate pass@k for these cases. To evaluate with other k values, pass --k=<comma-separated-values-here>. For other options, see

$ evaluate_infilling_functional_correctness --help

However, we recommend that you use the default values for the rest.

Known Issues

While evaluation uses very little memory, you might see the following error message when the system is running out of RAM. Since this may cause some correct programs to fail, we recommend that you free some memory and try again.

malloc: can't allocate region