kiranpantha/w2v-bert-2.0-nepali-ft-parliament

Wave2Vec2-Bert2.0 Nepali - Kiran Pantha

This model is a fine-tuned version of facebook/w2v-bert-2.0-nepali on the NepaliParliamentDS dataset. It achieves the following results on the evaluation set:

• Loss: 1.2860
• Wer: 0.6172
• Cer: 0.2217

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: 5e-05
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• optimizer: Use OptimizerNames.SGD and the args are: No additional optimizer arguments
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• num_epochs: 5
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Cer	Validation Loss	Wer
1.1608	0.3641	300	0.2240	1.3237	0.6323
1.1726	0.7282	600	0.2236	1.3175	0.6307
1.3014	1.0922	900	0.2231	1.3112	0.6284
1.1958	1.4563	1200	0.2229	1.3076	0.6276
1.1507	1.8204	1500	0.2228	1.3058	0.6269
1.1777	2.1845	1800	0.2227	1.3030	0.6253
1.1428	2.5485	2100	0.2223	1.2989	0.6235
1.2121	2.9126	2400	0.2222	1.2953	0.6220
1.221	3.2767	2700	1.2920	0.6207	0.2222
1.1533	3.6408	3000	1.2897	0.6198	0.2222
1.1758	4.0049	3300	1.2878	0.6185	0.2220
1.1918	4.3689	3600	1.2867	0.6177	0.2219
1.1421	4.7330	3900	1.2860	0.6172	0.2217

Framework versions

• Transformers 4.47.1
• Pytorch 2.6.0+xpu
• Datasets 3.2.0
• Tokenizers 0.21.0