cdli-whisper-en-ug-ke-sunbird-full-a40

This model is a fine-tuned version of Sunbird/asr-whisper-large-v3-salt on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.8134
• Wer: 0.2654
• Cer: 0.1807
• test_cer = 0.1119
• test_loss = 0.6422
• test_runtime = 0:22:41.66
• test_samples_per_second = 1.424
• test_steps_per_second = 0.356
• test_wer = 0.1834 #Both datasets combined

Ugandan English

• utterance_avg_wer = 0.215460
• utterance_avg_cer = 0.132733
• wer = 0.236585
• cer = 0.140362

Kenyan English

• utterance_avg_wer = 0.147041
• utterance_avg_cer = 0.088261
• wer = 0.146343
• cer = 0.087723

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: 4
• eval_batch_size: 4
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 16
• 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: 150
• training_steps: 3000

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
0.7169	0.4248	250	0.8477	0.2903	0.1997
0.6641	0.8496	500	0.8096	0.2821	0.1963
0.547	1.2736	750	0.8075	0.2777	0.1892
0.5109	1.6984	1000	0.7924	0.2720	0.1875
0.4612	2.1223	1250	0.7972	0.2656	0.1796
0.4589	2.5472	1500	0.8042	0.2666	0.1818
0.5135	2.9720	1750	0.7934	0.2710	0.1859
0.445	3.3959	2000	0.8043	0.2657	0.1813
0.4636	3.8207	2250	0.8070	0.2677	0.1831
0.371	4.2447	2500	0.8112	0.2655	0.1812
0.4372	4.6695	2750	0.8133	0.2672	0.1824
0.3861	5.0935	3000	0.8134	0.2654	0.1807

Framework versions

• Transformers 4.52.0
• Pytorch 2.7.1+cu118
• Datasets 3.6.0
• Tokenizers 0.21.4