refactor-benchmark VS llama-cpp-python

FWIW, I agree with you that each model has its own personality and that models may do better or worse on different kinds of coding tasks. Aider leans into both of these concepts.

The GPT-4 Turbo models have a lazy coding personality, and I spent months of effort figuring out how to both measure and reduce that laziness. This resulted in aider supporting "unified diffs" as a code editing format to reduce such laziness by 3X [0] and the aider refactoring benchmark as a way to quantify these benefits [1].

The benchmark results I just shared about GPT-4 Turbo with Vision cover both smaller, toy coding problems [2] as well as larger edits to larger source files [3]. The new model slightly underperforms on smaller coding tasks, and significantly underperforms on the larger edits where laziness is often a culprit.

[0] https://aider.chat/2023/12/21/unified-diffs.html

[1] https://github.com/paul-gauthier/refactor-benchmark

[2] https://aider.chat/2024/04/09/gpt-4-turbo.html#code-editing-...

[3] https://aider.chat/2024/04/09/gpt-4-turbo.html#lazy-coding

1-2 sentences: Rather than writing code, GPT-4 Turbo often inserts comments like "... finish implementing function here ...". I made a benchmark that provokes and quantifies that behavior.

1-2 paragraphs:

I found that I could provoke lazy coding by giving GPT-4 Turbo refactoring tasks, where I ask it to refactor a large method out of a large class. I analyzed 9 popular open source python repos and found 89 such methods that were conceptually easy to refactor, and built them into a benchmark [0].

GPT succeeds on a task if it can remove the method from its original class and add it to the top level of the file with appropriate changes to the SIZE of the abstract syntax tree. By measuring the size of the AST, we infer that GPT didn't replace a bunch of code with a comment like "... insert original method here...". I also gathered other laziness metrics like counting the number of new comments that contained "...", which correlated well with the AST size test.

[0] https://github.com/paul-gauthier/refactor-benchmark

There's a Python binding for llama.cpp which is actively maintained and has worked well for me: https://github.com/abetlen/llama-cpp-python

I'll share the core bit that took a while to figure out the right format, my main script is a hot mess using embeddings with SentenceTransformer, so I won't share that yet. E.g: last night I did a PR for llama-cpp-python that shows how Phi might be used with JSON only for the author to write almost exactly the same code at pretty much the same time. https://github.com/abetlen/llama-cpp-python/pull/1184

Python Bindings for llama.cpp

If the model could be made to work with llama.cpp, then https://github.com/abetlen/llama-cpp-python might be more compact. llama.cpp only supports a limited list of model types though.

https://github.com/abetlen/llama-cpp-python has a web server mode that replicates openai's API iirc and the readme shows it has docker builds already.

LocalAI https://localai.io/ and LMStudio https://lmstudio.ai/ both have fairly complete OpenAI compatibility layers. llama-cpp-python has a FastAPI server as well: https://github.com/abetlen/llama-cpp-python/blob/main/llama_... (as of this moment it hasn't merged GGUF update yet though)

I went with Python, llama-cpp-python, since my goal is just to get a small project up and running locally.