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.
6 comments
[ 2.5 ms ] story [ 30.5 ms ] threadIt’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...
https://www.reddit.com/r/LocalLLaMA/comments/1dhx449/deepsee...
links to two GGUFs:
https://huggingface.co/bartowski/DeepSeek-Coder-V2-Lite-Inst...
https://huggingface.co/LoneStriker/DeepSeek-Coder-V2-Lite-In...
Apparently both require disabling flash attention.
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.