nllb-1.3B-many-to-many-pronouncorrection-charaug

This model is a fine-tuned version of jq/nllb-1.3B-many-to-many-step-2k on the generator dataset. It achieves the following results on the evaluation set:

• Loss: 1.2075
• Bleu Ach Eng: 28.371
• Bleu Lgg Eng: 30.45
• Bleu Lug Eng: 41.978
• Bleu Nyn Eng: 32.296
• Bleu Teo Eng: 30.422
• Bleu Eng Ach: 20.972
• Bleu Eng Lgg: 22.362
• Bleu Eng Lug: 30.359
• Bleu Eng Nyn: 15.305
• Bleu Eng Teo: 21.391
• Bleu Mean: 27.391

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: 0.0003
• train_batch_size: 25
• eval_batch_size: 25
• seed: 42
• gradient_accumulation_steps: 120
• total_train_batch_size: 3000
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• training_steps: 1500
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Bleu Ach Eng	Bleu Lgg Eng	Bleu Lug Eng	Bleu Nyn Eng	Bleu Teo Eng	Bleu Eng Ach	Bleu Eng Lgg	Bleu Eng Lug	Bleu Eng Nyn	Bleu Eng Teo	Bleu Mean
No log	0.0667	100	1.1541	29.033	31.47	41.596	34.169	32.442	19.677	19.657	27.889	14.554	19.143	26.963
No log	1.0301	200	1.1570	27.473	31.853	41.934	32.575	31.606	20.25	20.634	28.592	13.672	19.997	26.859
No log	1.0968	300	1.1288	29.086	33.257	43.387	33.678	33.579	20.377	20.91	28.906	14.992	21.013	27.919
No log	2.0603	400	1.1620	28.122	31.46	42.491	33.304	32.331	20.282	21.604	29.577	14.961	20.94	27.507
0.7273	3.0237	500	1.1661	28.311	32.122	42.825	32.333	32.415	19.799	22.287	29.558	15.708	21.948	27.731
0.7273	3.0904	600	1.1652	28.593	30.62	41.964	33.383	32.08	21.142	21.8	30.215	14.717	21.744	27.626
0.7273	4.0538	700	1.2075	28.371	30.45	41.978	32.296	30.422	20.972	22.362	30.359	15.305	21.391	27.391

Framework versions

• Transformers 4.40.1
• Pytorch 2.2.0
• Datasets 2.19.0
• Tokenizers 0.19.1