Makkoen/whisper-large-v3-cit-do01-wd0-lr3e-06-steps2000-FULL6

FULL6

This model is a fine-tuned version of openai/whisper-large-v3 on the FULL-2024-11-22 dataset. It achieves the following results on the evaluation set:

• Loss: 0.3933
• Wer Ortho: 21.7759
• Wer: 15.7318

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: 3e-06
• train_batch_size: 4
• eval_batch_size: 8
• seed: 42
• distributed_type: multi-GPU
• gradient_accumulation_steps: 4
• total_train_batch_size: 16
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 300
• training_steps: 2000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer Ortho	Wer
0.688	0.3661	200	0.4873	26.5992	19.7836
0.5254	0.7323	400	0.4390	24.6602	18.1855
0.4648	1.0984	600	0.4158	22.9719	16.9557
0.4014	1.4645	800	0.4072	23.2981	17.1182
0.3921	1.8307	1000	0.3984	22.3407	16.2132
0.3684	2.1968	1200	0.3965	22.2350	16.3119
0.3326	2.5629	1400	0.3936	21.8665	15.6564
0.3331	2.9291	1600	0.3921	21.5282	15.4852
0.3032	3.2952	1800	0.3921	21.9390	15.8565
0.3007	3.6613	2000	0.3933	21.7759	15.7318

Framework versions

• Transformers 4.44.0
• Pytorch 1.13.1+cu117
• Datasets 2.21.0
• Tokenizers 0.19.1