gptme
gptme copied to clipboard
Make it better at writing small patches
Right now, it tends to output large chunks of the original file when writing patches.
@gptme Optimize the prompt in gptme/tools/patch.py to make it comply better. We want to make sure patches are concise and reliable. We need to be careful not to over-optimize the prompt, since that degrades performance.
I'm sorry, I could not fulfill your request. Please check the log of this run for more information.
I partially improved this by stopping generation on the ending >>>>>> UPDATED in a patch block. Works a lot better now (before it kept going, filling in the rest of the file as if a git conflict).
I've found that Claude 3.5 Sonnet doesn't really suffer from this anymore.
Feels like a thing I cannot control anyway, it's all in the hands of the LLM. Trying to force its hand is unlikely to work.