mms-1b-ckt-pooled-all-dev

This model is a fine-tuned version of facebook/mms-1b-all on a private Chukchi dataset.

It achieves the following results on the evaluation set (containing radio speech from Radio "Purga" and expedition speech from project Chuklang):

• Wer: 0.7275
• Cer: 0.1649

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: 0.001
• train_batch_size: 16
• eval_batch_size: 8
• 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: 25
• num_epochs: 12
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
0.6492	0.3205	50	0.6441	0.5978	0.1394
0.663	0.6410	100	0.6072	0.5780	0.1351
0.6973	0.9615	150	0.5798	0.5803	0.1321
0.6367	1.2821	200	0.5628	0.5487	0.1228
0.56	1.6026	250	0.5550	0.5504	0.1237
0.5971	1.9231	300	0.5333	0.5385	0.1160
0.619	2.2436	350	0.5332	0.5306	0.1138
0.4919	2.5641	400	0.5123	0.5171	0.1101
0.5333	2.8846	450	0.4979	0.5030	0.1063
0.5752	3.2051	500	0.5091	0.5135	0.1071
0.5634	3.5256	550	0.5067	0.5003	0.1067
0.4938	3.8462	600	0.4875	0.4970	0.1052
0.495	4.1667	650	0.4686	0.4858	0.1004
0.5397	4.4872	700	0.4673	0.4888	0.1002
0.4292	4.8077	750	0.4793	0.4875	0.1015
0.4992	5.1282	800	0.4577	0.4664	0.0965
0.4599	5.4487	850	0.4627	0.4786	0.0987
0.5029	5.7692	900	0.4566	0.4809	0.0976
0.4593	6.0897	950	0.4516	0.4674	0.0960
0.4309	6.4103	1000	0.4494	0.4691	0.0949
0.5021	6.7308	1050	0.4480	0.4717	0.0952
0.4574	7.0513	1100	0.4545	0.4615	0.0952
0.4127	7.3718	1150	0.4471	0.4589	0.0917
0.44	7.6923	1200	0.4450	0.4579	0.0935
0.5049	8.0128	1250	0.4405	0.4575	0.0937
0.3912	8.3333	1300	0.4472	0.4496	0.0925
0.4054	8.6538	1350	0.4373	0.4457	0.0932
0.5151	8.9744	1400	0.4397	0.4559	0.0927
0.4367	9.2949	1450	0.4409	0.4533	0.0924

Framework versions

• Transformers 4.57.1
• Pytorch 2.12.0+cu130
• Datasets 5.0.0
• Tokenizers 0.22.2