T5 base but trained at FP16 in the MRQA sample dataset. This model is the checkpoint at 3000 steps (3rd epoch), because there were instabilities during the late epochs.

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 3e-05
train_batch_size: 6
eval_batch_size: 6
seed: 42
gradient_accumulation_steps: 3
total_train_batch_size: 18
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: linear
num_epochs: 3 (5) (we take model checkpoint at 3rd epoch)
mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss
0.7978	0.9996	833	0.6668
0.6516	1.9992	1666	0.6532
0.6275	3.0	2500	0.6532
(0.6443)	(3.9996)	(3333)	(0.6533)
(2.0743)	(4.998)	(4165	(nan)

Note that this model is the checkpoint at 3000 steps (3rd epoch), because there were instabilities during the late epochs.

Framework versions

Transformers 4.42.0.dev0
Pytorch 2.3.0+cu121
Datasets 2.19.2
Tokenizers 0.19.1

enriquesaou/t5-base-mrqa-16

t5-base-mrqa-16

Model description