File size: 1,087 Bytes
0df2042
 
 
 
 
 
 
 
 
 
 
 
6732015
0df2042
 
6732015
 
 
 
 
 
 
 
 
 
 
 
 
0df2042
 
 
 
 
6732015
 
0df2042
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
{
  "base_model": "KPF/KPF-bert-ner",
  "training_dataset": "townboy/korean-pii-dataset",
  "training_dataset_rows": 11732,
  "splits": {
    "train": 9227,
    "validation": 1510,
    "test": 995
  },
  "pii_types": 33,
  "bio_labels": 67,
  "maximum_sequence_length": 512,
  "final_refinement": {
    "effective_batch_size": 32,
    "class_weighted_loss": false,
    "stages": [
      {
        "purpose": "Korean-ID counterfactual learning",
        "epochs": 1,
        "learning_rate": 5e-06
      },
      {
        "purpose": "general-label stabilization",
        "epochs": 1,
        "learning_rate": 1e-06
      }
    ],
    "validation_micro_f1": 0.9964634318856981
  },
  "final_evaluation": {
    "documents": 1320,
    "positive_spans_per_label": 100,
    "exact_character_span_matching": true,
    "macro_f1": 0.883369786048345,
    "micro_f1": 0.8642728407042839
  },
  "notes": "The final refinement used standard cross-entropy. The trained parameters are stored in model.safetensors; class weighting is a loss setting, not a separate required inference weight file."
}