pszemraj/pegasus-large-summary-explain

This model is a fine-tuned version of google/pegasus-large on the booksum dataset for four total epochs.

It achieves the following results on the evaluation set:

• eval_loss: 1.1193
• eval_runtime: 6.6754
• eval_samples_per_second: 27.714
• eval_steps_per_second: 1.798
• epoch: 3.0
• step: 900

A 1-epoch checkpoint can be found at pszemraj/pegasus-large-book-summary , which is where the second training session started from.

Model description

• After some initial tests, it was found that models trained on the booksum dataset seem to inherit the summaries' SparkNotes-style explanations; so the user gets a shorter and easier-to-understand version of the text instead of just more compact.
• This quality (anecdotally) is favourable for learning/comprehension because summarization datasets that simply make the information more compact (* cough * arXiv) can be so dense that the overall time spent trying to comprehend what it is saying can be the same as just reading the original material.

Intended uses & limitations

• standard pegasus has a max input length of 1024 tokens, therefore the model only saw the first 1024 tokens of a chapter when training, and learned to try to make the chapter's summary from that. Keep this in mind when using this model, as information at the end of a text sequence longer than 1024 tokens may be excluded from the final summary/the model will be biased towards information presented first.

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 4e-05
• train_batch_size: 16
• eval_batch_size: 16
• seed: 42
• distributed_type: multi-GPU
• gradient_accumulation_steps: 2
• total_train_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.03
• num_epochs: 4

Framework versions

• Transformers 4.16.2
• Pytorch 1.10.2+cu113
• Datasets 1.18.3
• Tokenizers 0.11.0