Help me write a poem: Instruction Tuning as a Vehicle for Collaborative Poetry Writing

Repo associated with the EMNLP 2022 paper (https://arxiv.org/pdf/2210.13669.pdf)

Code

• Scripts to train the model with both deepspeed and model parallelism are in training_scripts with an associated README.
• Once the models were trained we run automatic evaluation, comparing to pretrained models (T0 3B/ T0pp) and InstructGPT (via the OpenAI API), as directed by the README in eval_scripts/. We report success rates of models satisfying instructions, averaged over 5 runs.
• Subsequently, we also perform human evaluation comparing outputs from our finetuned T5-11B model and InstructGPT as seen in human_eval/.
• Finally we run a user study using the interface in demo_poetic/. Logs of all user interactions are provided in human_eval/poems_submissions

Data

• Drive Link to the training and validation data in the instruction format along with the test sets used in the paper.
• Model generations used in the paper for automatic evaluation.
• Poems and user interactions in the collaborative user study we ran with CoPoet.
• Model weights of our finetuned T5-11B model (TODO)

Citation

                              @article{chakrabarty2022help,
                                title={Help me write a poem: Instruction Tuning as a Vehicle for Collaborative Poetry Writing},
                                author={Chakrabarty, Tuhin and Padmakumar, Vishakh and He, He},
                                journal={arXiv preprint arXiv:2210.13669},
                                year={2022}
                              }