arxiv27k-t5-abst-title-gen/

This model is a fine-tuned version of mt5-small on the arxiv-abstract-title dataset. It achieves the following results on the evaluation set: - Loss: 1.6002 - Rouge1: 32.8 - Rouge2: 21.9 - Rougel: 34.8

Model description

Model has been trained with a colab-pro notebook in 4 hours.

Intended uses & limitations

Can be used for generating journal titles from given abstracts

Training args

model_args = T5Args() model_args.max_seq_length = 256 model_args.train_batch_size = 8 model_args.eval_batch_size = 8 model_args.num_train_epochs = 6 model_args.evaluate_during_training = False model_args.use_multiprocessing = False model_args.fp16 = False model_args.save_steps = 40000 model_args.save_eval_checkpoints = False model_args.save_model_every_epoch = True model_args.output_dir = OUTPUT_DIR model_args.no_cache = True model_args.reprocess_input_data = True model_args.overwrite_output_dir = True model_args.num_return_sequences = 1

Framework versions

• Transformers 4.12.5
• Pytorch 1.10.0+cu111
• Datasets 1.15.1
• Tokenizers 0.10.3

Contact

detasar@gmail.com Davut Emre Taşar