Fine-tuning codellama with own dataset for code review

Hi :) I am planning to fine tune Code LLaMA 7B using my own dataset to evaluate its effectiveness for code review tasks. I came across a paper where the authors attempted a similar approach using data from CodeReviewer (https://arxiv.org/abs/2203.09095) .

My plan is to gather data from the pull requests in our repositories and create JSON files in the following format: { "oldf": "... contents of another old file ...", "patch": "@@ -25,13 +25,16 @@ ...", // Diffs from the repo "msg": "we call cities + towns ...", "id": 12959, "y": 1 } Will this data format work for fine tuning Code LLaMA, or will I need to adjust the dataset to a specific format for it to be compatible? Thanks in advance