Which model do I use for my custom seq2seq data?

Hello HuggingFace community,

An example from my train data is as follows:

"Input": "
          User: Why is my peach tree not producing fruit? 
          AI: Lack of fruit can be due to poor pollination, insufficient chill hours, or nutrient imbalances. 
          User: How to ensure enough chill hours?
"Output": "User: How can I ensure my peach tree gets enough chill hours to produce fruit?"

Basically, the Output is the last question in the Input, modified to be more clear, adds the contextual information and resolves coreferences so when it stands alone it has all the information that the question in the Input meant. I need to train a model so that given the above Input a similar Output as shown above is generated. The model should be lightweight (<1GB), with fast inference time so as to make it scalable.

I generated around 10k such examples using GPT-3.5-turbo. I have tried finetuning the current data on “facebook\bart” from hugging face for 5 epochs, it performs decently but not so perfect to use it live.

I just randomly chose bart. Is there any other model which might be better for this kind of data? Do I need to generate more training examples?