Regional_BnASR

This model is a fine-tuned version of outputs/checkpoint-11500 on an unknown dataset. It achieves the following results on the evaluation set:

• Loss: 2.8498
• Wer: 0.7432

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: 2e-05
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 500
• training_steps: 20000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
2.4681	1.25	500	9.8862	0.9367
2.0154	2.51	1000	9.1645	0.9174
1.9749	3.76	1500	7.8777	0.8787
1.9293	5.02	2000	8.7211	0.8949
1.8588	6.27	2500	8.5633	0.8851
1.8637	7.52	3000	8.9099	0.8895
1.7904	8.79	3500	8.5780	0.8846
1.8647	10.04	4000	8.9881	0.8900
1.8229	11.3	4500	8.6372	0.8818
1.8365	12.56	5000	8.5890	0.8793
1.7912	13.81	5500	8.6685	0.8809
1.792	15.07	6000	8.6862	0.8796
1.7924	16.33	6500	8.5065	0.8763
1.7685	17.58	7000	8.8943	0.8840
1.8004	18.84	7500	9.0298	0.8861
1.7792	20.09	8000	8.8783	0.8818
1.7749	21.35	8500	8.8410	0.8811
1.8002	22.61	9000	8.8083	0.8804
1.7496	23.86	9500	8.8536	0.8815
1.7625	25.12	10000	8.8653	0.8816
1.7468	26.38	10500	8.8854	0.8826
1.7533	27.63	11000	8.8604	0.8799
1.7507	28.89	11500	8.6150	0.8743
1.766	30.14	12000	9.2126	0.8894
1.7641	31.39	12500	8.7537	0.8770
1.7673	32.65	13000	8.7769	0.8764
1.7281	33.92	13500	8.8294	0.8767
1.7315	35.17	14000	9.0875	0.8827
1.7271	36.42	14500	9.0619	0.8826
1.8015	52.26	15000	2.8379	0.7426
1.8405	54.0	15500	2.8486	0.7434
1.8206	55.73	16000	2.8677	0.7445
1.8401	57.47	16500	2.8111	0.7418
1.781	59.21	17000	2.8444	0.7430
1.7892	60.95	17500	2.8829	0.7445
1.8344	62.68	18000	2.8498	0.7432

Framework versions

• Transformers 4.37.0.dev0
• Pytorch 2.1.0+cu121
• Datasets 2.14.6
• Tokenizers 0.15.0