fix_punct_cased_t5_small

This model is a fine-tuned version of google/t5-v1_1-small on the NPR utterances dataset.

Dataset

The model was trained on 80k rows from the above dataset consisting of NPR radio transcripts. Commans, periods, and semicolons were removed from the text and then random commas, periods, and semicolons were added. The model was trained to place those three punctuation marks in the correct location. The casing of the texts was not modified during training.

It achieves the following results on the evaluation set:

• Loss: 0.2744
• Rouge1: 93.3712
• Rouge2: 91.0027
• Rougel: 93.3618
• Rougelsum: 93.3479
• Gen Len: 46.0728

Model description

The purpose of this model is to correct the punctuation in a sentence. For example, the phrase "This is, a sentence. with odd punctuation to show off what, the model. can do" gets changed to "This is a sentence with odd punctuation to show off what the model can do."

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 8e-05
• train_batch_size: 128
• eval_batch_size: 256
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 3.0

Training results

Training Loss	Epoch	Step	Validation Loss	Rouge1	Rouge2	Rougel	Rougelsum	Gen Len
1.2254	1.0	600	0.3501	63.2952	59.8766	63.137	63.2022	16.2637
0.7345	2.0	1200	0.2815	64.896	61.6256	64.8677	64.8728	16.3625
0.6536	3.0	1800	0.2744	64.8724	61.6282	64.8483	64.8502	16.3906

Framework versions

• Transformers 4.22.2
• Pytorch 1.11.0a0+17540c5
• Datasets 2.5.1
• Tokenizers 0.12.1