This is flan-t5-base finetuned on wookieepedia. First, the data was downloaded using: https://robvanderg.github.io/datasets/wikia/ , then I created a script for doing LM with T5, which is uploaded in this repository (run_t5_mlm_torch.py; note that not all functionality is tested). Finally, it was trained with the following command:
python3 retrain_torch.py --train_file ../starwarsfandomcom-20200223.txt.cleaned.tok.uniq.txt --output_dir flan-t5-base-starwars --validation_split_percentage 1 --model_name_or_path google/flan-t5-base --max_seq_length 512 --do_train --do_eval
Final perplexity was 3.38
- Downloads last month
- 1