[GH-ISSUE #230] Train deepseek on the compiler error annotations #87

New issue

Open

opened 2026-10-26 09:45:11 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-26 09:45:11 +00:00

Owner

Originally created by @hkalbasi on GitHub (Jun 18, 2026).
Original GitHub issue: https://github.com/deepseek-ai/awesome-deepseek-agent/issues/230

Sorry if it is offtopic here, had no contact way with deepseek team. Hope this works. Feel free to close it if the message is received and you don't want to track it here.

I'm using deepseek v4 flash, which is a great model for agentic coding overall. I found one problem which is probably easily fixable for next models. The problem is that deepseek doesn't understand terminal annotations of code used by compilers and linters, like:

this is foo
            ^^^ foo is here

you can give it tasks like annotating foo in the text, and almost always it makes an off by one error. But other frontier models seems are specifically trained for this and do these kind of tasks perfectly.

I think by doing a RL training on these annotations, like the task I mentioned, the understanding of the compile errors would improved massively, which is very important in agentic coding. I saw deepseek fighting with the compiler on off by one edit spans, just for lack of understanding the position of these annotations.

Hope it helps a little in making deepseek greater than closed source models.

Originally created by @hkalbasi on GitHub (Jun 18, 2026). Original GitHub issue: https://github.com/deepseek-ai/awesome-deepseek-agent/issues/230 Sorry if it is offtopic here, had no contact way with deepseek team. Hope this works. Feel free to close it if the message is received and you don't want to track it here. I'm using deepseek v4 flash, which is a great model for agentic coding overall. I found one problem which is probably easily fixable for next models. The problem is that deepseek doesn't understand terminal annotations of code used by compilers and linters, like: ``` this is foo ^^^ foo is here ``` you can give it tasks like annotating foo in the text, and almost always it makes an off by one error. But other frontier models seems are specifically trained for this and do these kind of tasks perfectly. I think by doing a RL training on these annotations, like the task I mentioned, the understanding of the compile errors would improved massively, which is very important in agentic coding. I saw deepseek fighting with the compiler on off by one edit spans, just for lack of understanding the position of these annotations. Hope it helps a little in making deepseek greater than closed source models.

bsakaguchi referenced this issue

2026-10-26 09:46:41 +00:00

[PR #87] Add Xpro: AI-powered desktop IDE with DeepSeek V4 support #194