gwanju_largeWER_model

This model is a fine-tuned version of openai/whisper-large on the Marcusxx/gwanju dataset. It achieves the following results on the evaluation set:

• Loss: 0.3334
• Wer: 41.8546

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: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 100
• training_steps: 4000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.4683	0.0741	250	0.4884	104.4328
0.4578	0.1482	500	0.4522	55.8304
0.4675	0.2223	750	0.4379	65.3948
0.4338	0.2964	1000	0.4225	65.4206
0.4547	0.3705	1250	0.4023	63.5814
0.3676	0.4446	1500	0.3914	47.9551
0.3752	0.5187	1750	0.3840	48.3838
0.3584	0.5928	2000	0.3745	44.8641
0.4221	0.6669	2250	0.3638	42.4548
0.3432	0.7410	2500	0.3563	42.7206
0.3993	0.8151	2750	0.3497	44.7955
0.3448	0.8892	3000	0.3437	43.3722
0.3441	0.9632	3250	0.3381	40.4270
0.2317	1.0373	3500	0.3350	39.5782
0.2063	1.1114	3750	0.3339	40.8385
0.2016	1.1855	4000	0.3334	41.8546

Framework versions

• Transformers 4.41.2
• Pytorch 2.2.2+cu121
• Datasets 2.19.2
• Tokenizers 0.19.1