t5-v1_1-base-ft-jflAUG

A more recent version can be found here. Training smaller and/or comparably sized models is a WIP.

t5-v1_1-base-ft-jflAUG

GOAL: a more robust and generalized grammar and spelling correction model that corrects everything in a single shot. It should have a minimal impact on the semantics of correct sentences (i.e. it does not change things that do not need to be changed).

• this model (at least from preliminary testing) can handle large amounts of errors in the source text (i.e. from audio transcription) and still produce cohesive results.
• a fine-tuned version of google/t5-v1_1-base on an expanded version of the JFLEG dataset.

Model description

• this is a WIP. This fine-tuned model is v1.
• long term: a generalized grammar and spelling correction model that can handle lots of things at the same time.
• currently, it seems to be more of a "gibberish to mostly correct English" translator

Intended uses & limitations

• try some tests with the examples here
• thus far, some limitations are: sentence fragments are not autocorrected (at least, if entered individually), some more complicated pronoun/they/he/her etc. agreement is not always fixed.

Training and evaluation data

• trained as text-to-text
• JFLEG dataset + additional selected and/or generated grammar corrections

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 6e-05
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• distributed_type: multi-GPU
• gradient_accumulation_steps: 8
• total_train_batch_size: 64
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.05
• num_epochs: 5

Framework versions

• Transformers 4.17.0
• Pytorch 1.10.0+cu111
• Datasets 2.0.0
• Tokenizers 0.11.6