Edit model card

xls-r-300m-dv

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: 0.2855
  • Wer: 0.2665

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: 0.0003
  • train_batch_size: 16
  • eval_batch_size: 8
  • seed: 42
  • gradient_accumulation_steps: 2
  • total_train_batch_size: 32
  • optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
  • lr_scheduler_type: linear
  • lr_scheduler_warmup_steps: 500
  • num_epochs: 50
  • mixed_precision_training: Native AMP

Training results

Training Loss Epoch Step Validation Loss Wer
4.3386 0.66 400 1.1411 0.9432
0.6543 1.33 800 0.5099 0.6749
0.4646 1.99 1200 0.4133 0.5968
0.3748 2.65 1600 0.3534 0.5515
0.3323 3.32 2000 0.3635 0.5527
0.3269 3.98 2400 0.3587 0.5423
0.2984 4.64 2800 0.3340 0.5073
0.2841 5.31 3200 0.3279 0.5004
0.2664 5.97 3600 0.3114 0.4845
0.2397 6.63 4000 0.3174 0.4920
0.2332 7.3 4400 0.3110 0.4911
0.2304 7.96 4800 0.3123 0.4785
0.2134 8.62 5200 0.2984 0.4557
0.2066 9.29 5600 0.3013 0.4723
0.1951 9.95 6000 0.2934 0.4487
0.1806 10.61 6400 0.2802 0.4547
0.1727 11.28 6800 0.2842 0.4333
0.1666 11.94 7200 0.2873 0.4272
0.1562 12.6 7600 0.3042 0.4373
0.1483 13.27 8000 0.3122 0.4313
0.1465 13.93 8400 0.2760 0.4226
0.1335 14.59 8800 0.3112 0.4243
0.1293 15.26 9200 0.3002 0.4133
0.1264 15.92 9600 0.2985 0.4145
0.1179 16.58 10000 0.2925 0.4012
0.1171 17.25 10400 0.3127 0.4012
0.1141 17.91 10800 0.2980 0.3908
0.108 18.57 11200 0.3108 0.3951
0.1045 19.24 11600 0.3269 0.3908
0.1047 19.9 12000 0.2998 0.3868
0.0937 20.56 12400 0.2918 0.3875
0.0949 21.23 12800 0.2906 0.3657
0.0879 21.89 13200 0.2974 0.3731
0.0854 22.55 13600 0.2943 0.3711
0.0851 23.22 14000 0.2919 0.3580
0.0789 23.88 14400 0.2983 0.3560
0.0796 24.54 14800 0.3131 0.3544
0.0761 25.21 15200 0.2996 0.3616
0.0755 25.87 15600 0.2972 0.3506
0.0726 26.53 16000 0.2902 0.3474
0.0707 27.2 16400 0.3083 0.3480
0.0669 27.86 16800 0.3035 0.3330
0.0637 28.52 17200 0.2963 0.3370
0.0596 29.19 17600 0.2830 0.3326
0.0583 29.85 18000 0.2969 0.3287
0.0566 30.51 18400 0.3002 0.3480
0.0574 31.18 18800 0.2916 0.3296
0.0536 31.84 19200 0.2933 0.3225
0.0548 32.5 19600 0.2900 0.3179
0.0506 33.17 20000 0.3073 0.3225
0.0511 33.83 20400 0.2925 0.3275
0.0483 34.49 20800 0.2919 0.3245
0.0456 35.16 21200 0.2859 0.3105
0.0445 35.82 21600 0.2864 0.3080
0.0437 36.48 22000 0.2989 0.3084
0.04 37.15 22400 0.2887 0.3060
0.0406 37.81 22800 0.2870 0.3013
0.0397 38.47 23200 0.2793 0.3020
0.0383 39.14 23600 0.2955 0.2943
0.0345 39.8 24000 0.2813 0.2905
0.0331 40.46 24400 0.2845 0.2845
0.0338 41.13 24800 0.2832 0.2925
0.0333 41.79 25200 0.2889 0.2849
0.0325 42.45 25600 0.2808 0.2847
0.0314 43.12 26000 0.2867 0.2801
0.0288 43.78 26400 0.2865 0.2834
0.0291 44.44 26800 0.2863 0.2806
0.0269 45.11 27200 0.2941 0.2736
0.0275 45.77 27600 0.2897 0.2736
0.0271 46.43 28000 0.2857 0.2695
0.0251 47.1 28400 0.2881 0.2702
0.0243 47.76 28800 0.2901 0.2684
0.0244 48.42 29200 0.2849 0.2679
0.0232 49.09 29600 0.2849 0.2677
0.0224 49.75 30000 0.2855 0.2665

Framework versions

  • Transformers 4.17.0.dev0
  • Pytorch 1.10.2+cu102
  • Datasets 1.18.3
  • Tokenizers 0.11.0
Downloads last month
14

Dataset used to train shahukareem/xls-r-300m-dv

Evaluation results