What's the preferred file format for LLMs editing data files? JSON5/yaml/etc? Making a [apply_patch](https://github.com/anomalyco/opencode/blob/dev/packages/opencode/src/tool/apply_patch.txt) tool call and curious what file format I should provide for best accuracy across model providers

So this is a tricky one to answer. I like to assume (or at least think I know..) that the pre-training data set is very rich in JSON knowledge and YAML knowledge Based on the trend set by OpenAI early on - JSON has become a defacto in strucutred output, tool calling methods, and the way instruction tuning has become a standard its also being used to help teach LLM via RLHF to product JSON ... So my "off the cuff" answer is JSON. that being said TOON/Yaml/CSV/TSv are all protocols and "serialized" formats that are well understood, declarative and pretty easy for the bigger models to handle. But if you looking at the 7B and under they primarily are focused on the JSON landscape... so for that I would still lean towards it. --- Now to answer form the perspective of what "file type is easiest to apply a patch too" .. kinda subjective here.. How nested is your data file - if its super nested.. well JSON really is the king (TOON is a cool play on it.. more compressed representation, but harder to read as a human/model .. = parsing via code can be a fun time) If its not super nested and more flat then a CSV is by far the easier way ^^ This is also highly depending on how you plan to parse using your tool.. Do you want to do a json.loads, or a csv.loads ... both are the same effort in code. But JSON has lots more in terms of nested flexibility