SPEAK-ASR/whisper-si-exp-4

SPEAK-ASR/whisper-small-si-openslr_plus_youtube_data

This model is a fine-tuned version of openai/whisper-small on the Whisper Small - Sinhala ASR Fine-Tuned dataset. It achieves the following results on the evaluation set:

• Loss: 0.1465

Model description

• WandB ID: hardy-hill-51
• Used full OpenSLR dataset + Our Youtube dataset
• On eval didn't use Youtube dataset test split
• Used LoRA adapters
• Previous name was SPEAK-ASR/whisper-small-si-openslr_plus_youtube_data

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: 64
• eval_batch_size: 32
• seed: 42
• 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: 500
• num_epochs: 3

Training results

Training Loss	Epoch	Step	Validation Loss
0.3137	0.4254	1000	0.2489
0.2126	0.8507	2000	0.1943
0.1506	1.2761	3000	0.1737
0.1835	1.7014	4000	0.1607
0.1879	2.1268	5000	0.1538
0.1485	2.5521	6000	0.1494
0.1473	2.9775	7000	0.1465

Framework versions

• PEFT 0.18.1
• Transformers 5.0.0
• Pytorch 2.10.0+cu128
• Datasets 4.5.0
• Tokenizers 0.22.2