shan-stt-v3

This model is a fine-tuned version of k1ngtai/shan-stt-v2 on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.2857
• Model Preparation Time: 0.0053
• Wer: 89.5425
• Cer: 19.6965

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: 16
• eval_batch_size: 8
• seed: 42
• 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: linear
• lr_scheduler_warmup_steps: 100
• training_steps: 2000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Model Preparation Time	Wer	Cer
0.0042	12.5	250	0.2420	0.0053	92.1569	23.2369
0.0002	25.0	500	0.2624	0.0053	88.8889	20.4552
0.0000	37.5	750	0.2754	0.0053	90.1961	20.0056
0.0000	50.0	1000	0.2817	0.0053	90.1961	19.8932
0.0000	62.5	1250	0.2857	0.0053	89.5425	19.6965
0.0000	75.0	1500	0.2886	0.0053	89.5425	19.9494
0.0000	87.5	1750	0.2907	0.0053	89.5425	20.0899
0.0000	100.0	2000	0.2912	0.0053	90.1961	20.0056

Framework versions

• Transformers 5.0.0
• Pytorch 2.10.0+cu128
• Datasets 3.6.0
• Tokenizers 0.22.2