distilhubert-ko-zeroth

This model is a fine-tuned version of ntu-spml/distilhubert on the BINGSU/ZEROTH-KOREAN - NA dataset. It achieves the following results on the evaluation set:

• Loss: 0.9934
• Cer: 0.2066

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: 0.001
• 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: cosine_with_restarts
• lr_scheduler_warmup_steps: 500
• num_epochs: 10.0
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Cer
No log	0.57	400	3.2681	0.6761
7.285	1.15	800	1.5312	0.4170
1.259	1.72	1200	1.3459	0.3846
0.9108	2.3	1600	1.1357	0.3239
0.7227	2.87	2000	1.0571	0.3056
0.7227	3.45	2400	1.0002	0.2829
0.5689	4.02	2800	0.8773	0.2553
0.4676	4.6	3200	0.8634	0.2462
0.3805	5.17	3600	0.8504	0.2323
0.2548	5.75	4000	0.8480	0.2260
0.2548	6.32	4400	0.8550	0.2231
0.189	6.9	4800	0.8587	0.2159
0.1336	7.47	5200	0.9012	0.2101
0.0827	8.05	5600	0.9302	0.2100
0.0506	8.62	6000	0.9622	0.2063
0.0506	9.2	6400	0.9826	0.2062
0.0389	9.77	6800	0.9933	0.2067

Framework versions

• Transformers 4.21.1
• Pytorch 1.12.1
• Datasets 2.4.0
• Tokenizers 0.12.1