T5-base-finetuned-mnli

This model is T5 fine-tuned on GLUE MNLI dataset. It acheives the following results on the validation-matched set

• Accuracy: 0.8567

Model Details

T5 is an encoder-decoder model pre-trained on a multi-task mixture of unsupervised and supervised tasks and for which each task is converted into a text-to-text format.

Training procedure

Tokenization

Since, T5 is a text-to-text model, the labels of the dataset are converted as follows: For each example, a sentence as been formed as "mnli premise: " + mnli_premise + "hypothesis: " + mnli_hypothesis and fed to the tokenizer to get the input_ids and attention_mask. For each label, target is choosen as "entailment" if label is 0, else it is "neutral" if label is 1, else it is "contradiction" and tokenized to get input_ids and attention_mask . During training, these inputs_ids having pad token are replaced with -100 so that loss is not calculated for them. Then these input ids are given as labels, and above attention_mask of labels is given as decoder attention mask.

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 3e-4
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• optimizer: epsilon=1e-08
• num_epochs: 2

Training results

Epoch	Training Loss	Validation Matched Accuracy
1	0.1661	0.8404
2	0.1016	0.8567