Edit model card

wav2vec2-large-xls-r-300m-kinyarwanda

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the common_voice dataset. It achieves the following results on the evaluation set:

  • Loss: 0.3917
  • Wer: 0.3246

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: 7e-05
  • train_batch_size: 12
  • eval_batch_size: 8
  • seed: 42
  • gradient_accumulation_steps: 2
  • total_train_batch_size: 24
  • optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
  • lr_scheduler_type: linear
  • lr_scheduler_warmup_steps: 400
  • num_epochs: 8
  • mixed_precision_training: Native AMP

Training results

Training Loss Epoch Step Validation Loss Wer
9.0634 0.12 400 3.0554 1.0
2.8009 0.24 800 1.5927 0.9554
0.9022 0.36 1200 0.7328 0.6445
0.6213 0.48 1600 0.6138 0.5510
0.5299 0.6 2000 0.6072 0.5223
0.4999 0.72 2400 0.5449 0.4969
0.4731 0.84 2800 0.5261 0.4828
0.458 0.96 3200 0.5058 0.4607
0.4158 1.09 3600 0.4892 0.4463
0.4037 1.21 4000 0.4759 0.4429
0.4021 1.33 4400 0.4615 0.4330
0.3934 1.45 4800 0.4593 0.4315
0.3808 1.57 5200 0.4736 0.4344
0.3838 1.69 5600 0.4569 0.4249
0.3726 1.81 6000 0.4473 0.4140
0.3623 1.93 6400 0.4403 0.4097
0.3517 2.05 6800 0.4389 0.4061
0.333 2.17 7200 0.4383 0.4104
0.3354 2.29 7600 0.4360 0.3955
0.3257 2.41 8000 0.4226 0.3942
0.3275 2.53 8400 0.4206 0.4040
0.3262 2.65 8800 0.4172 0.3875
0.3206 2.77 9200 0.4209 0.3877
0.323 2.89 9600 0.4177 0.3825
0.3099 3.01 10000 0.4101 0.3691
0.3008 3.14 10400 0.4055 0.3709
0.2918 3.26 10800 0.4085 0.3800
0.292 3.38 11200 0.4089 0.3713
0.292 3.5 11600 0.4092 0.3730
0.2785 3.62 12000 0.4151 0.3687
0.2941 3.74 12400 0.4004 0.3639
0.2838 3.86 12800 0.4108 0.3703
0.2854 3.98 13200 0.3911 0.3596
0.2683 4.1 13600 0.3944 0.3575
0.2647 4.22 14000 0.3836 0.3538
0.2704 4.34 14400 0.4006 0.3540
0.2664 4.46 14800 0.3974 0.3553
0.2662 4.58 15200 0.3890 0.3470
0.2615 4.7 15600 0.3856 0.3507
0.2553 4.82 16000 0.3814 0.3497
0.2587 4.94 16400 0.3837 0.3440
0.2522 5.06 16800 0.3834 0.3486
0.2451 5.19 17200 0.3897 0.3414
0.2423 5.31 17600 0.3864 0.3481
0.2434 5.43 18000 0.3808 0.3416
0.2525 5.55 18400 0.3795 0.3408
0.2427 5.67 18800 0.3841 0.3411
0.2411 5.79 19200 0.3804 0.3366
0.2404 5.91 19600 0.3800 0.3328
0.2372 6.03 20000 0.3749 0.3335
0.2244 6.15 20400 0.3820 0.3327
0.2381 6.27 20800 0.3789 0.3325
0.2294 6.39 21200 0.3867 0.3298
0.2378 6.51 21600 0.3843 0.3281
0.2312 6.63 22000 0.3813 0.3277
0.2411 6.75 22400 0.3780 0.3268
0.2315 6.87 22800 0.3790 0.3280
0.241 6.99 23200 0.3776 0.3281
0.2313 7.11 23600 0.3929 0.3283
0.2423 7.24 24000 0.3905 0.3280
0.2337 7.36 24400 0.3979 0.3249
0.2368 7.48 24800 0.3980 0.3257
0.2409 7.6 25200 0.3937 0.3229
0.2416 7.72 25600 0.3867 0.3237
0.2364 7.84 26000 0.3912 0.3253
0.234 7.96 26400 0.3917 0.3246

Framework versions

  • Transformers 4.11.3
  • Pytorch 1.10.0+cu113
  • Datasets 1.18.3
  • Tokenizers 0.10.3
Downloads last month
2
Hosted inference API
or or
This model can be loaded on the Inference API on-demand.

Dataset used to train peter2000/wav2vec2-large-xls-r-300m-kinyarwanda