MathildeB3/52Hz-small-fr-v3

52Hz Small Fr - IMT Atlantique X 52 Hertz

This model is a fine-tuned version of openai/whisper-small on the Transcriptions IMTx52Hz v3 dataset. It achieves the following results on the evaluation set:

• Loss: 0.4065
• Wer: 17.8105

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: 4
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 16
• 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: cosine
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 15
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
1.4611	1.0	22	0.8947	39.8693
0.6451	2.0	44	0.5211	25.4902
0.3082	3.0	66	0.4803	20.4248
0.224	4.0	88	0.4255	37.7451
0.1531	5.0	110	0.4062	17.9739
0.1294	6.0	132	0.4146	19.1176
0.1018	7.0	154	0.4123	19.6078
0.0546	8.0	176	0.4034	20.2614
0.0401	9.0	198	0.4386	17.3203
0.0476	10.0	220	0.4225	18.6275
0.0261	11.0	242	0.4131	18.4641
0.0226	12.0	264	0.4035	17.8105
0.0323	13.0	286	0.4050	17.6471
0.0208	14.0	308	0.4064	17.8105
0.0189	15.0	330	0.4065	17.8105

Framework versions

• Transformers 4.57.3
• Pytorch 2.9.1+cu130
• Datasets 4.4.2
• Tokenizers 0.22.2