bart-base-detox

This model is a fine-tuned version of facebook/bart-base on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.1819

Model description

More information needed

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: 1e-05
• train_batch_size: 4
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 16
• total_train_batch_size: 64
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 20

Training results

Training Loss	Epoch	Step	Validation Loss
0.5633	1.0	135	0.2524
0.2589	2.0	270	0.2193
0.2307	3.0	405	0.1993
0.2171	4.0	540	0.2002
0.2027	5.0	675	0.1937
0.1946	6.0	810	0.1972
0.1874	7.0	945	0.1917
0.1853	8.0	1080	0.1868
0.1811	9.0	1215	0.1890
0.1776	10.0	1350	0.1871
0.1798	11.0	1485	0.1858
0.1745	12.0	1620	0.1820
0.1689	13.0	1755	0.1827
0.1707	14.0	1890	0.1843
0.1658	15.0	2025	0.1834
0.1647	16.0	2160	0.1820
0.1645	17.0	2295	0.1837
0.1633	18.0	2430	0.1814
0.1612	19.0	2565	0.1815
0.1603	20.0	2700	0.1819

Framework versions

• Transformers 4.16.2
• Pytorch 1.12.0.dev20220429
• Datasets 2.1.0
• Tokenizers 0.10.3