This model is a fine-tuned version of facebook/mms-1b-all on a private dataset, intended to achieve the best quality specifically on Chukchi expedition data from project Chuklang.

It achieves the following results on the evaluation set (evaluation set containing only expedition Chuklang data):

Wer: 0.7391
Cer: 0.1825

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: 16
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: 24
mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
1.2212	1.1905	50	0.8237	0.8397	0.2052
0.9554	2.3810	100	0.7470	0.7778	0.1877
0.8978	3.5714	150	0.7269	0.8	0.1921
0.8594	4.7619	200	0.7058	0.7587	0.1820
0.7897	5.9524	250	0.7039	0.7413	0.1769
0.7547	7.1429	300	0.6790	0.7111	0.1701
0.7233	8.3333	350	0.6716	0.7016	0.1677
0.6611	9.5238	400	0.6658	0.7190	0.1712
0.6446	10.7143	450	0.6566	0.6952	0.1637
0.6057	11.9048	500	0.6674	0.6937	0.1610
0.6135	13.0952	550	0.6605	0.6714	0.1584
0.5837	14.2857	600	0.6455	0.7079	0.1615
0.5391	15.4762	650	0.6601	0.6937	0.1619
0.5051	16.6667	700	0.6519	0.6889	0.1606

Framework versions

Transformers 4.57.1
Pytorch 2.11.0+cu128
Datasets 4.8.5
Tokenizers 0.22.2

tadgeis/mms-1b-all-ckt-best-cer-model

mms-1b-ckt-staged-bible-radio-to-chuklang-source-eval-stage2

Model description

Intended uses & limitations

Training and evaluation data

Training procedure

Training hyperparameters

Training results

Framework versions

tadgeis/
mms-1b-all-ckt-best-cer-model