Marialab/finetuned-whisper-small-dr-ar

Finetuned Whisper small darija translate

This model is a fine-tuned version of openai/whisper-small on the Darija-C dataset. It achieves the following results on the evaluation set:

• Loss: 0.0000
• Bleu: 0.7440

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: 1
• eval_batch_size: 1
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 4
• optimizer: Use 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_steps: 40
• training_steps: 200
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Bleu
3.975	0.6667	10	3.4363	0.0
2.1029	1.3333	20	1.5986	0.0262
1.7909	2.0	30	0.9239	0.1314
0.9837	2.6667	40	0.5086	0.3289
0.36	3.3333	50	0.4370	0.3911
0.6361	4.0	60	0.2622	0.4561
0.5227	4.6667	70	0.2506	0.5266
0.3307	5.3333	80	0.1299	0.6123
0.2438	6.0	90	0.1290	0.6057
0.2864	6.6667	100	0.0838	0.6623
0.073	7.3333	110	0.0965	0.6494
0.1924	8.0	120	0.0859	0.7237
0.0086	8.6667	130	0.0235	0.7174
0.003	9.3333	140	0.0335	0.7354
0.054	10.0	150	0.0096	0.7367
0.0008	10.6667	160	0.0190	0.7367
0.0002	11.3333	170	0.0003	0.7440
0.0001	12.0	180	0.0001	0.7440
0.0002	12.6667	190	0.0001	0.7440
0.0	13.3333	200	0.0000	0.7440

Framework versions

• Transformers 4.46.3
• Pytorch 2.5.1+cu121
• Datasets 2.19.2
• Tokenizers 0.20.3