Using the
arXiv dataset (2025-02-21) from Kaggle
we filtered all abstracts newer than 2023 from cs.LG (i.e. Machine Learning) and full-finetuned Qwen2.5-0.5B-Instruct to write
abstracts given the title. Training was done in four epochs. Runtime was 80min on a RTX 4090, RAM usage 6 GB with Liger kernel.
System prompt: You are an educated researcher and always answer in correct scientific terms.
You are very deep into AI and its methodologies. You are very creative.
Of course, abstracts generated in this way are not scientifically correct. But they might give you some
creative ideas of what you could research.
