ntnu-smil/sandi-exp-closed-full-merged

sandi-exp-closed-full

This model is a fine-tuned version of openai/whisper-large-v3 on the ntnu-smil/sandi2025-ds dataset. It achieves the following results on the evaluation set:

• Loss: 1.1903
• Wer: 50.5791
• Cer: 132.7118
• Decode Runtime: 258.3093
• Wer Runtime: 0.2561
• Cer Runtime: 0.4723

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: 8
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 128
• total_train_batch_size: 1024
• optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.98) and epsilon=1e-06 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• training_steps: 56

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer	Decode Runtime	Wer Runtime	Cer Runtime
2.5401	1.1435	7	1.4718	41.7132	115.4428	266.2705	0.3015	0.6137
1.4085	2.2870	14	1.4074	43.0855	119.3428	252.7248	0.2317	0.5350
1.3715	3.4305	21	1.3522	45.6971	124.1181	255.2491	0.2517	0.4809
1.3325	4.5740	28	1.2988	47.7710	127.5660	254.4386	0.2534	0.4564
1.2652	5.7175	35	1.2547	49.3496	130.5721	256.8112	0.2712	0.4814
1.2605	6.8610	42	1.2205	50.1876	132.1325	263.2403	0.2855	0.5078
2.0534	8.1435	49	1.1986	50.5748	132.4652	259.0558	0.2459	0.4612
1.2319	9.2870	56	1.1903	50.5791	132.7118	258.3093	0.2561	0.4723

Framework versions

• PEFT 0.15.2
• Transformers 4.48.2
• Pytorch 2.4.1+cu124
• Datasets 3.5.1
• Tokenizers 0.21.1