whisper-small-karelian-cs-w-rus

This model is a fine-tuned version of openai/whisper-small on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.7827
• Wer: 0.4393
• Cer: 0.1569

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: 4
• eval_batch_size: 4
• seed: 42
• distributed_type: multi-GPU
• gradient_accumulation_steps: 8
• total_train_batch_size: 32
• optimizer: Use OptimizerNames.ADAMW_BNB 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: 10000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
0.4436	0.4813	500	0.7520	0.5127	0.1571
0.2601	0.9627	1000	0.6816	0.4746	0.1602
0.2059	1.4438	1500	0.7251	0.4717	0.1724
0.1789	1.9252	2000	0.6968	0.4551	0.1674
0.1262	2.4063	2500	0.6831	0.4475	0.1415
0.0987	2.8876	3000	0.6758	0.4227	0.1353
0.0693	3.3687	3500	0.7007	0.4426	0.1382
0.059	3.8501	4000	0.6895	0.4195	0.1354
0.0429	4.3312	4500	0.6934	0.4276	0.1565
0.0411	4.8125	5000	0.7100	0.4270	0.1336
0.0414	5.2936	5500	0.7505	0.4455	0.1430
0.0424	5.7750	6000	0.7456	0.4683	0.1723
0.0407	6.2561	6500	0.7290	0.4372	0.1452
0.034	6.7374	7000	0.7417	0.4290	0.1361
0.0334	7.2185	7500	0.7427	0.4310	0.1410
0.028	7.6999	8000	0.7573	0.4149	0.1343
0.0216	8.1810	8500	0.7471	0.4185	0.1368
0.0197	8.6623	9000	0.7490	0.4220	0.1383
0.0244	9.1434	9500	0.7671	0.4301	0.1423
0.0219	9.6248	10000	0.7827	0.4393	0.1569

Framework versions

• Transformers 4.51.3
• Pytorch 2.7.0+cu126
• Datasets 3.5.1
• Tokenizers 0.21.1