The main problem of conditional text generation is that it is mainly based on the content of an input set of examples: this leads to little diversification of the generated text. To overcome this shortcoming, we have fine tuned CTRL using three different datasets. The first model has been used as a baseline for comparison, while the other two have been used to obtain more formal and informal text. The BART model has been employed for text classification to gauge formality.
I have encountered problems to prepare the dataset. As the information from the README file, the original dataset could be downloaded directly from COCO or kaggle. However, when I am tracing into the split_data.py file, I found the lines 5 and 6 are the original input of the train_validate and test npy files. I wanna ask you here is that how did you prepare these npy files. Is that directly download from the COCO? or did you do additional preprocessing after getting the original train, valid and test files from the COCO website?