Training Loss	Epoch	Step	Validation Loss	Wer
2.9114	0.29	1000	inf	0.9997
1.2436	0.57	2000	inf	0.4310
1.0552	0.86	3000	inf	0.3144
1.0044	1.15	4000	inf	0.2814
0.9718	1.43	5000	inf	0.2658
0.9502	1.72	6000	inf	0.2566
0.9418	2.01	7000	inf	0.2476
0.9215	2.29	8000	inf	0.2420
0.9236	2.58	9000	inf	0.2388
0.9014	2.87	10000	inf	0.2354
0.8814	3.15	11000	inf	0.2312
0.8809	3.44	12000	inf	0.2285
0.8717	3.73	13000	inf	0.2263
0.8787	4.01	14000	inf	0.2218
0.8567	4.3	15000	inf	0.2193
0.8488	4.59	16000	inf	0.2187
0.8359	4.87	17000	inf	0.2172

Training continued with checkpoint from STEP 17000:

Training Loss	Epoch	Step	Validation Loss	Wer
/	5.16	18000	inf	0.2176
/	5.45	19000	inf	0.2181
/	5.73	20000	inf	0.2155
/	6.02	21000	inf	0.2140
/	6.31	22000	inf	0.2124
/	6.59	23000	inf	0.2117
/	6.88	24000	inf	0.2116

It achieves the best result on the validation set on Step 24000:

Wer: 0.2116

Got some issue with validation loss calculation.

Framework versions

Transformers 4.17.0.dev0
Pytorch 1.10.2+cu102
Datasets 1.18.3.dev0
Tokenizers 0.11.0

Evaluation Commands

To evaluate on mozilla-foundation/common_voice_8 with split test

python eval.py --model_id Plim/xls-r-300m-cv_8-fr --dataset mozilla-foundation/common_voice_8_0 --config fr --split test

To evaluate on speech-recognition-community-v2/dev_data

python eval.py --model_id Plim/xls-r-300m-cv_8-fr --dataset speech-recognition-community-v2/dev_data --config fr --split validation --chunk_length_s 5.0 --stride_length_s 1.0