mT5_tai-lo_to_chinese_LoRA_ver1.0.k

This model is a fine-tuned version of google/mt5-base on an unknown dataset. It achieves the following results on the evaluation set:

• Loss: 2.1796
• Chrf: 16.1759

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: 0.001
• train_batch_size: 8
• eval_batch_size: 16
• seed: 1
• optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: constant_with_warmup
• lr_scheduler_warmup_steps: 4000
• training_steps: 10000

Training results

Training Loss	Epoch	Step	Validation Loss	Chrf
9.1427	0.9337	1000	3.4432	0.1313
3.6246	1.8674	2000	3.2296	0.713
3.4624	2.8011	3000	3.3720	0.2161
3.2719	3.7348	4000	2.9802	2.1503
3.1149	4.6685	5000	2.8530	3.3285
2.8582	5.6022	6000	2.5916	7.6784
2.6203	6.5359	7000	2.4117	7.1325
2.4322	7.4697	8000	2.3129	12.998
2.2501	8.4034	9000	2.2295	11.9602
2.1502	9.3371	10000	2.1796	16.1759

Framework versions

• PEFT 0.19.1
• Transformers 5.0.0
• Pytorch 2.10.0+cu128
• Datasets 4.0.0
• Tokenizers 0.22.2