asr-africa/mms-300m_DigitalUmuganda_Afrivoice_Shona_20hr_v1

facebook/mms-300m

This model is a fine-tuned version of facebook/mms-300m on the DigitalUmuganda dataset. It achieves the following results on the evaluation set:

• Loss: 0.7396
• Wer: 0.4128
• Cer: 0.0953

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: 0.0003
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 32
• 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: 100
• num_epochs: 150
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
23.8246	0.9954	109	2.9644	1.0	1.0
11.6666	2.0	219	2.8975	1.0	1.0
11.5956	2.9954	328	2.8467	1.0	1.0
10.3096	4.0	438	1.9758	1.0	0.6253
5.5486	4.9954	547	0.7431	0.9025	0.2232
2.263	6.0	657	0.3758	0.5049	0.0977
1.4016	6.9954	766	0.3167	0.4258	0.0805
1.0646	8.0	876	0.2821	0.3744	0.0702
0.8883	8.9954	985	0.2754	0.3581	0.0679
0.7279	10.0	1095	0.2653	0.3543	0.0669
0.644	10.9954	1204	0.2632	0.3417	0.0610
0.5561	12.0	1314	0.2647	0.3336	0.0597
0.5029	12.9954	1423	0.2868	0.3368	0.0614
0.434	14.0	1533	0.2981	0.3239	0.0583
0.3766	14.9954	1642	0.2986	0.3327	0.0604
0.3356	16.0	1752	0.3170	0.3150	0.0562
0.3046	16.9954	1861	0.3350	0.3278	0.0579
0.2801	18.0	1971	0.3359	0.3358	0.0598
0.2601	18.9954	2080	0.3392	0.3341	0.0599
0.2339	20.0	2190	0.3526	0.3282	0.0584
0.2175	20.9954	2299	0.3517	0.3317	0.0582
0.2031	22.0	2409	0.3547	0.3275	0.0564
0.197	22.9954	2518	0.3688	0.3164	0.0552
0.1751	24.0	2628	0.3724	0.3212	0.0570
0.1738	24.9954	2737	0.3778	0.3035	0.0539
0.1656	26.0	2847	0.3946	0.3120	0.0547
0.163	26.9954	2956	0.3770	0.3089	0.0542
0.1484	28.0	3066	0.3871	0.3151	0.0568
0.1452	28.9954	3175	0.3712	0.3038	0.0535
0.1335	30.0	3285	0.3989	0.3085	0.0544
0.1337	30.9954	3394	0.3880	0.3064	0.0543
0.126	32.0	3504	0.4123	0.3079	0.0540
0.1252	32.9954	3613	0.3959	0.3070	0.0544
0.1229	34.0	3723	0.3975	0.2989	0.0533
0.1198	34.9954	3832	0.4194	0.3020	0.0531
0.1136	36.0	3942	0.3916	0.3056	0.0534
0.1186	36.9954	4051	0.3939	0.3040	0.0535
0.1078	38.0	4161	0.4042	0.3011	0.0532
0.108	38.9954	4270	0.4044	0.3000	0.0526
0.1011	40.0	4380	0.4309	0.2997	0.0522
0.106	40.9954	4489	0.4245	0.2941	0.0517
0.0994	42.0	4599	0.4268	0.3059	0.0528
0.0982	42.9954	4708	0.4239	0.2996	0.0524
0.0956	44.0	4818	0.4175	0.2974	0.0525
0.0931	44.9954	4927	0.4677	0.2934	0.0518
0.0886	46.0	5037	0.4262	0.2979	0.0523
0.0874	46.9954	5146	0.4274	0.2927	0.0514
0.0882	48.0	5256	0.4230	0.2988	0.0524
0.0856	48.9954	5365	0.4169	0.2872	0.0505
0.0826	50.0	5475	0.4092	0.3034	0.0522
0.0828	50.9954	5584	0.4383	0.2869	0.0507
0.0798	52.0	5694	0.4201	0.2968	0.0512
0.0761	52.9954	5803	0.3999	0.2871	0.0498
0.0753	54.0	5913	0.4303	0.2914	0.0502
0.0743	54.9954	6022	0.4424	0.2937	0.0507
0.0788	56.0	6132	0.4214	0.2877	0.0495
0.0752	56.9954	6241	0.4293	0.2925	0.0507
0.0713	58.0	6351	0.4494	0.2919	0.0504
0.0693	58.9954	6460	0.4297	0.2882	0.0498

Framework versions

• Transformers 4.47.0.dev0
• Pytorch 2.1.0+cu118
• Datasets 3.0.2
• Tokenizers 0.20.1