aider/docs/benchmarks.md

GPT code editing benchmarks

Aider is an open source command line chat tool that lets you ask GPT to edit code in your local git repos. You can use aider to ask GPT to add features, write tests or make other changes and improvements to your code.

Having a reliable way for GPT to edit local source code files is critical to providing this functionality. Making code editing more reliable often involves changing and experimenting with the "edit format" that aider uses. The edit format is a key part of the system prompt, specifying how GPT should format code edits in its replies. Different edit formats can range in complexity from something simple like "return an updated copy of the whole file" to a much more sophisticated format that uses OpenAI's new function calling API to specify a series of specific diffs

To measure the impact of changes to the edit format, I created a benchmark based on the Exercism python coding exercises. This benchmark measures how well aider & GPT can turn a natural language coding request into actual runnable code saved into files that pass unit tests. This is an end-to-end assessment of not just how well GPT can write code, but also how well it can edit existing code and package up those code changes so that aider can save the edits to the local source files.

I ran this code editing benchmark on almost all the ChatGPT models, using a variety of edit formats. This produced some interesting results:

• Asking GPT to return an updated copy of the whole file in a normal markdown fenced code block is by far the most reliable and effective edit format. This is true across all GPT-3.5 and GPT-4 models.
• Using the new function calling API is worse than the above whole file method, for all models. GPT writes worse code and frequently mangles this output format, even though the function calling API was introduced to make structured outputs more reliable. This was a big surprise.
• The new June (0613) versions of gpt-3.5-turbo are worse at code editing than the older February (0301) version. This was unexpected.
• The GPT-4 models are much better at code editing than the GPT-3.5 models, as expected.

The quantitative benchmark results agree with an intuition that I've been developing about how to prompt GPT for complex tasks like coding. You want to minimize the "cognitive overhead" of formatting the response, so that GPT can focus on the task at hand. As an analogy, you wouldn't expect a good result if you asked a junior developer to implement a new feature by hand typing the required code changes as diff -c formatted edits.

Using more complex output formats seems to cause two problems:

• It makes GPT write worse code. Keeping the output format simple seems to leave GPT with more attention to devote to the actual coding task.
• It makes GPT less likely to adhere to the output format. This makes it harder for tooling to correctly identify and apply the edits it is trying to make.

I had hoped that the new function calling API would enable more reliable use of structured output formats, but it does not appear to be a panacea when working with source code.

More details on the benchmark, edit formats and results are discussed below.

The benchmark

The benchmark uses 133 practice exercises from the Exercism python repository. These exercises were designed for people to learn python and practice their coding skills.

Each exercise has:

• Some instructions for the exercise, in markdown files.
• Stub code for the implementation in a python file, specifying the functions/classes that need to be implemented.
• Unit tests in a seperate python file.

The goal is for GPT to read the instructions, implement the provided functions/class skeletons and pass all the unit tests. The benchmark measures what percentage of the 133 exercises are completed successfully, causing all the associated unit tests to pass.

To complete an exercise, aider sends GPT the initial contents of the implementation file, the Exercism instructions and a final instruction:

Use the above instructions to modify the supplied files: <implementation file>
Keep and implement the existing function or class stubs, they will be called from unit tests.
Only use standard python libraries, don't suggest installing any packages.

Aider updates the implementation file based on GPT's reply and runs the unit tests. If they all pass, we are done. If some tests fail, aider sends GPT a second message with the test error output. It only sends the first 50 lines of test errors, to avoid exhausting the context window of the smaller models. Aider also includes this final instruction:

See the testing errors above.
The tests are correct.
Fix the code in <implementation file> to resolve the errors.

Requiring GPT to fix its first implementation in response to test failures is another way in which this benchmark stresses code editing skill. This second chance is also important because it gives a chance for GPT to adjust if the instructions were imprecise with respect to the specific requirements of the unit tests. Many of the exercises have multiple paragraphs of instructions, and most human coders would likely fail some tests on their first try.

It's worth noting that GPT never gets to see the source code of the unit tests. Just the error output from failed tests.

In summary, passing an exercise means GPT was able to:

• write the required code (possibly after reviewing test error output),
• correctly package up all of this code into the edit format so that aider can process and save it to the implementation file.

Conversely, failing an exercise only requires a breakdown in one of those steps. In practice, GPT fails at different steps in different exercises. Sometimes it just writes the wrong code. Other times, it fails to format the code edits in a way that conforms to the edit format so the code isn't saved properly.

It's worth keeping in mind that changing the edit format often affects both aspects of GPT's performance. Complex edit formats often make it write worse code and make it less successful at formatting the edits correctly.

Edit formats

I benchmarked 4 different edit formats, described below. Each description includes a sample response that GPT might provide in response to a user who requests: "Change the print from hello to goodbye."

whole

The whole format asks GPT to return an updated copy of the entire file, including any changes. The file should be formatted with normal markdown triple-backtick fences, inlined with the rest of its response text.

This format is very similar to how ChatGPT returns code snippets during normal chats, except with the addition of a filename right before the opening triple-backticks.

Here is the updated copy of your file demo.py:

demo.py
```python
def main():
    print("goodbye")
```

diff

The diff format also asks GPT to return edits as part of the normal response text, in a simple diff format. Each edit is a fenced code block that specifies the filename and a chunk of ORIGINAL and UPDATED code. GPT provides some original lines from the file and then a new updated set of lines.

Here are the changes you requested to demo.py:

```python
demo.py
<<<<<<< ORIGINAL
    print("hello")
=======
    print("goodbye")
>>>>>>> UPDATED
```

whole-func

The whole-func format requests updated copies of whole files to be returned using the function call API.

{
    "explanation": "Changed hello to goodbye.",
    "files": [
        {
            "path": "demo.py",
            "content": "def main():\n    print(\"goodbye\")\n"
        }
}

diff-func

The diff-func format requests a list of original/updated style edits to be returned using the function call API.

{
    "explanation": "Changed hello to goodbye.",
    "edits": [
        {
            "path": "demo.py",
            "original_lines": [
                "    print(\"hello\")"
            ],
            "updated_lines": [
                "    print(\"goodbye\")"
            ],
        }
    ]
}       

GPT-3.5 struggles with complex edit formats

While GPT-3.5 is able to pass some exercises using edit formats other than the whole format, it really struggles with the rest of the formats.

Pathlogical use of diff

While GPT-3.5 is sometimes able to correctly generate the diff edit format, it often uses it in a pathological way.

It places the entire original source file in the ORIGINAL block and the entire updated file in the UPDATED block. This is strictly worse than just using the whole edit format, since GPT is sending 2 full copies of the file.

Hallucinating function calls

When using the functions API GPT-3.5 is prone to ignoring the JSON Schema that specifies valid functions. It often returns a completely novel and semantically invalid function_call fragment with "name": "python".

        "function_call": {
          "name": "python",
          "arguments": "def main():\n    print(\"hello\")\n"
        },

The arguments attribute is supposed to be a set of key/value pairs with the arguments to the function specified in the name field. Instead, GPT-3.5 frequently just stuffs an entire python file into that field.

It feels like it might be getting confused by fine tuning that was done for the ChatGPT coder interpreter plugin?

Randomness

The benchmark attempts to be deterministic, always sending identical requests for each exercise on repeated runs. As part of this effort, when sending test error output to GPT it removes the wall-clock timing information that is normally included by the unittest module.

The benchmarking harness also logs sha hashes of all the OpenAI API requests and replies. This makes it possible to detect randomness or nondeterminism in the bechmarking process.

It turns out that the OpenAI chat APIs are not deterministic, even at temperature=0. The same identical request will produce multiple distinct responses, usually on the order of 3-6 different variations. This feels like OpenAI may be load balancing their API across a number of slightly different instances of the model.

For some exercises, some of these variable responses pass the unit tests while other variants do not. Results for exercises like this which are "on the bubble" are therefore a bit random, depending on which variant OpenAI returns.

Given that, it would be ideal to run all 133 exercises many times for each model/edit-format combination and report an average performance. This would average away the effect of the API variance. It would also significantly increase the cost of this sort of benchmarking. So I didn't do that.

Benchmarking against 133 exercises provides some robustness all by itself, since we are measuring the performance across many exercises.

But to get a sense of how much the API variance impacts the benchmark outcomes, I ran all 133 exercises 10 times each against gpt-3.5-turbo-0613 with the whole edit format. You'll see one set of error bars in the graph, which show the range of results from those 10 runs.

The OpenAI API randomness doesn't seem to cause a large variance in the benchmark results.

Conclusions

Based on these benchmarking results, aider will continue to use the whole edit format for GPT-3.5, and diff for GPT-4. While GPT-4 gets somewhat better results with the whole edit format, it significantly increases costs and latency compared to diff.

The latency of streaming back the entire updated copy of each edited file is the real challenge. The GPT-3.5 models are quite responsive, and can stream back entire files at an acceptable speed. Aider displays a progress bar and live diffs of the files as they stream in, which helps pass the time.

The GPT-4 models are much slower, and waiting for even small files to be completely "retyped" on each request is probably unacceptable.