opus-mt-cantonese-v2

This model is a fine-tuned version of Helsinki-NLP/opus-mt-en-zh. It achieves the following results on the evaluation set:

• Loss: 1.8948
• Bleu: 4.1931
• Gen Len: 12.2403

Model description

This model translates English into Cantonese.

Intended uses & limitations

Translations produced are for experimental purposes. Correctness is not guaranteed. Use at your own risk.

Training and evaluation data

Trained with Cantonese sentences with English translations:

• 6280 from Tatoeba.
• 1232 from CantoDict.

Training procedure

80% training / 20% validation. 20 epochs.

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 16
• eval_batch_size: 16
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 20
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Bleu	Gen Len
No log	1.0	376	2.2222	1.4485	11.5939
2.5401	2.0	752	1.9855	1.9434	11.6465
1.7798	3.0	1128	1.8805	1.1681	11.9527
1.4633	4.0	1504	1.8257	4.4967	12.6944
1.4633	5.0	1880	1.7930	4.3143	12.4161
1.1973	6.0	2256	1.8002	4.3407	12.0053
1.0388	7.0	2632	1.7965	5.4944	12.4075
0.9027	8.0	3008	1.7912	4.1324	12.486
0.9027	9.0	3384	1.8020	4.5473	12.2044
0.7774	10.0	3760	1.8120	4.3802	12.2237
0.6928	11.0	4136	1.8218	5.683	12.484
0.6229	12.0	4512	1.8346	4.8839	12.229
0.6229	13.0	4888	1.8521	5.6593	12.2157
0.5645	14.0	5264	1.8634	4.2758	12.1684
0.5114	15.0	5640	1.8679	4.3432	12.2463
0.4809	16.0	6016	1.8734	4.7905	12.2224
0.4809	17.0	6392	1.8869	5.4113	12.2184
0.447	18.0	6768	1.8902	5.4278	12.2543
0.4278	19.0	7144	1.8943	4.2225	12.4095
0.4179	20.0	7520	1.8948	4.1931	12.2403

Framework versions

• Transformers 4.39.2
• Pytorch 2.2.1+cu121
• Datasets 2.18.0
• Tokenizers 0.15.2