nrlt/whisper-large-v2-mn-2

Whisper Large V2 - Mongolian LoRA

This model is a fine-tuned version of openai/whisper-large-v2 on the nrlt/dataset-mn-preprocessed dataset. It achieves the following results on the evaluation set:

• Loss: 0.3712
• Wer: 38.7426

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: 0.0001
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 3
• total_train_batch_size: 24
• optimizer: Use OptimizerNames.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_ratio: 0.1
• num_epochs: 15

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
1.4376	1.4151	500	0.6133	65.9333
0.4635	2.8302	1000	0.4080	49.1834
0.3072	4.2453	1500	0.3462	44.3583
0.228	5.6604	2000	0.3182	40.4206
0.1724	7.0755	2500	0.3149	38.7426
0.1246	8.4906	3000	0.3212	39.0260
0.0963	9.9057	3500	0.3326	42.9935
0.0679	11.3208	4000	0.3589	39.4287
0.0512	12.7358	4500	0.3712	38.7426

Framework versions

• PEFT 0.19.1
• Transformers 4.46.3
• Pytorch 2.4.1+cu124
• Datasets 2.18.0
• Tokenizers 0.20.3