The new Deepseek Coder V2 model [0] is now atop aider’s code editing leaderboard!
It’s worth noting that it is only capable of using aider’s “whole” edit format. This means it returns a modified full copy of each file when it makes changes. Most other strong models are able to use aider’s “diff” editing format, which allows them to return diffs of edits – saving time and token costs and avoiding limits on output tokens.
You know i get it ... ML is primarily python... but why can't we get code models tuned for languages used by non-ML developers more targeted, .NET, PHP, Rust, C, as first class training datasets.
Shit with rusts great feedback from the compiler theres really no reason we can't have a full loop LLM trained on the code and the responses to properly handle the full dev cycle.
It’s worth noting that it is only capable of using aider’s “whole” edit format. This means it returns a modified full copy of each file when it makes changes. Most other strong models are able to use aider’s “diff” editing format, which allows them to return diffs of edits – saving time and token costs and avoiding limits on output tokens.
[0] https://huggingface.co/deepseek-ai/DeepSeek-Coder-V2-Instruc...