Full code and details at https://github.com/csinva/gpt-paper-title-generator

Model

• finetunes starting from thegpt-neo-2.7B checkpoint
  • for training details see the training script
• inference
  • should prepend with a year and two newlines before querying for a title, e.g. 2022\n\n

from transformers import AutoModelForCausalLM, pipeline, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("csinva/gpt-neo-2.7B-titles")
tokenizer = AutoTokenizer.from_pretrained("EleutherAI/gpt-neo-2.7B")
pipe = pipeline('text-generation', model=model, tokenizer=tokenizer)
pipe('2022\n\n')

Data

• all papers on arXiv in the categories cs.AI, cs.LG, stat.ML
  • date cutoff: only finetuned on papers with dat on or before Apr 1, 2022
  • random 5% of papers also excluded
  • this results in 98,388 papers for finetuning
• during finetuning each paper title was given starting with the prompt <year>\n\n <title>\n (e.g. 2022\n\n Emb-GAM: an Interpretable and Efficient Predictor using Pre-trained Language Models\n)