Fine tuning chatgpt giving poor results

I am trying to fine-tune davinci model to generate summary of a blob of text with emphasis on specific entities (Basically the summary I want is different from what is normally returned by the model when prompted). I prepared ~200 examples in jsonl format with prompt key containing the blob of text and completion key containing the summary I want. I also included a separator at the end of prompt as well as added a END token in completion key.

However the results are not what I want. Many times the output is not focussing on the entities of interest. Sometimes the result is worse than the normal model’s output.

Any tips on what can be done to improve fine tuning is appreciated.