w2v-bert-2.0-bengali-colab-CV17.0-v2

This model is a fine-tuned version of facebook/w2v-bert-2.0 on the common_voice_17_0 dataset. It achieves the following results on the evaluation set:

• Loss: 2.9399
• Wer: 1.0015

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: 2
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 8
• 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: linear
• lr_scheduler_warmup_steps: 50
• num_epochs: 5
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
No log	0.16	50	8.6700	1.3648
7.4868	0.32	100	6.2568	2.1074
7.4868	0.48	150	4.6350	1.0779
4.3743	0.64	200	3.6517	1.0015
4.3743	0.8	250	3.4099	1.0002
3.3984	0.96	300	3.3160	1.0004
3.3984	1.1184	350	3.2545	1.0004
3.2053	1.2784	400	3.2081	1.0004
3.2053	1.4384	450	3.1663	1.0007
3.1315	1.5984	500	3.1441	1.0004
3.1315	1.7584	550	3.1187	1.0004
3.071	1.9184	600	3.1005	1.0004
3.071	2.0768	650	3.0800	1.0009
3.0241	2.2368	700	3.0587	1.0007
3.0241	2.3968	750	3.0362	1.0007
2.9979	2.5568	800	3.0259	1.0007
2.9979	2.7168	850	3.0151	1.0009
2.9738	2.8768	900	3.0034	1.0009
2.9738	3.0352	950	2.9906	1.0007
2.9602	3.1952	1000	2.9799	1.0002
2.9602	3.3552	1050	2.9766	1.0007
2.9325	3.5152	1100	2.9698	1.0004
2.9325	3.6752	1150	2.9635	1.0004
2.8989	3.8352	1200	2.9588	1.0009
2.8989	3.9952	1250	2.9507	1.0007
2.9026	4.1536	1300	2.9501	1.0011
2.9026	4.3136	1350	2.9466	1.0007
2.8872	4.4736	1400	2.9435	1.0011
2.8872	4.6336	1450	2.9413	1.0009
2.9015	4.7936	1500	2.9405	1.0011
2.9015	4.9536	1550	2.9399	1.0015

Framework versions

• Transformers 4.56.2
• Pytorch 2.8.0+cu128
• Datasets 3.6.0
• Tokenizers 0.22.1