vumichien/ner-jp-gliner

This model is a fine-tuned version of deberta-v3-base-japanese on the Japanese Ner Wikipedia dataset. It achieves the following results:

• Precision: 96.07%
• Recall: 89.16%
• F1 score: 92.49%

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:

• num_steps: 30000
• train_batch_size: 8
• eval_every: 3000
• warmup_ratio: 0.1
• scheduler_type: "cosine"
• loss_alpha: -1
• loss_gamma: 0
• label_smoothing: 0
• loss_reduction: "sum"
• lr_encoder: 1e-5
• lr_others: 5e-5
• weight_decay_encoder: 0.01
• weight_decay_other: 0.01

Training results

Epoch	Training Loss
1	1291.582200
2	53.290100
3	44.137400
4	35.286200
5	20.865500
6	15.890000
7	13.369600
8	11.599500
9	9.773400
10	8.372600
11	7.256200
12	6.521800
13	7.203800
14	7.032900
15	6.189700
16	6.897400
17	6.031700
18	5.329600
19	5.411300
20	5.253800
21	4.522000
22	5.107700
23	4.163300
24	4.185400
25	3.403100
26	3.272400
27	2.387800
28	3.039400
29	2.383000
30	1.895300
31	1.748700
32	1.864300
33	2.343000
34	1.356600
35	1.182000
36	0.894700
37	0.954900