whisper-medium-bigcgen-combined-20hrs-model

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

• Loss: 0.5368
• Wer: 0.4211

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: 1.75e-05
• train_batch_size: 2
• eval_batch_size: 2
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 8
• optimizer: Use adamw_torch 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: 500
• training_steps: 5000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
3.7679	0.1521	200	0.8942	0.6455
3.4733	0.3042	400	0.7609	0.5650
2.6698	0.4564	600	0.6958	0.5375
2.5647	0.6085	800	0.6685	0.5496
2.1636	0.7606	1000	0.6228	0.5103
2.7265	0.9127	1200	0.5869	0.4706
1.5404	1.0654	1400	0.5990	0.4542
1.9844	1.2175	1600	0.5893	0.4643
1.6926	1.3697	1800	0.5730	0.4413
1.8654	1.5218	2000	0.5550	0.4599
1.8045	1.6739	2200	0.5445	0.4178
1.8258	1.8260	2400	0.5368	0.4211
1.543	1.9781	2600	0.5371	0.4245
0.9667	2.1308	2800	0.5587	0.4545
1.0216	2.2829	3000	0.5617	0.4096

Framework versions

• Transformers 4.47.1
• Pytorch 2.5.1+cu124
• Datasets 3.2.0
• Tokenizers 0.21.0