{'eval_loss': 0.10526859760284424,
 'eval_precision': 0.8299675891298928,
 'eval_recall': 0.8870237143618439,
 'eval_f1': 0.8575476558475013,
 'eval_accuracy': 0.9717195641875889,
 'eval_runtime': 18.2172,
 'eval_samples_per_second': 80.967,
 'eval_steps_per_second': 5.105,
 'epoch': 10.0}

Test from the Test set

{'eval_loss': 0.11170374602079391,
 'eval_precision': 0.8178285159096429,
 'eval_recall': 0.8823375262054507,
 'eval_f1': 0.8488591957645278,
 'eval_accuracy': 0.968742017138478,
 'eval_runtime': 18.6202,
 'eval_samples_per_second': 79.054,
 'eval_steps_per_second': 4.941,
 'epoch': 10.0}

How to use

Inference

Huggingface doesn't support inference token classification for Thai and It will give wrong tag. You must using this code.

python

from transformers import AutoTokenizer
from transformers import AutoModelForTokenClassification
from pythainlp.tokenize import word_tokenize # pip install pythainlp
import torch

name="Porameht/wangchanberta-thainer-corpus-v2-2"
tokenizer = AutoTokenizer.from_pretrained(name)
model = AutoModelForTokenClassification.from_pretrained(name)

sentence="นายปรเมศ คุ้มสมบัติ 552/44 หมู่ 1 บ้านหนองบัว ต.ภูหอ อ.ภูหลวง จ.เลย 42230"
cut=word_tokenize(sentence.replace(" ", "<_>"))
inputs=tokenizer(cut,is_split_into_words=True,return_tensors="pt")

ids = inputs["input_ids"]
mask = inputs["attention_mask"]
# forward pass
outputs = model(ids, attention_mask=mask)
logits = outputs[0]

predictions = torch.argmax(logits, dim=2)
predicted_token_class = [model.config.id2label[t.item()] for t in predictions[0]]

def fix_span_error(words,ner):
    _ner = []
    _ner=ner
    _new_tag=[]
    for i,j in zip(words,_ner):
        #print(i,j)
        i=tokenizer.decode(i)
        if i.isspace() and j.startswith("B-"):
            j="O"
        if i=='' or i=='<s>' or i=='</s>':
            continue
        if i=="<_>":
            i=" "
        _new_tag.append((i,j))
    return _new_tag

ner_tag=fix_span_error(inputs['input_ids'][0],predicted_token_class)
ner_tag

output:

[('นาย', 'B-PERSON'),
 ('ปร', 'I-PERSON'),
 ('เม', 'I-PERSON'),
 ('ศ', 'I-PERSON'),
 (' ', 'B-LOCATION'),
 ('คุ้ม', 'I-PERSON'),
 ('สมบัติ', 'I-PERSON'),
 (' ', 'O'),
 ('55', 'O'),
 ('2/', 'O'),
 ('44', 'O'),
 (' ', 'B-LOCATION'),
 ('หมู่', 'B-LOCATION'),
 (' ', 'I-LOCATION'),
 ('1', 'I-LOCATION'),
 (' ', 'B-LOCATION'),
 ('บ้าน', 'B-LOCATION'),
 ('หนอง', 'I-LOCATION'),
 ('บัว', 'I-LOCATION'),
 (' ', 'B-LOCATION'),
 ('ต', 'B-LOCATION'),
 ('.', 'I-LOCATION'),
 ('ภู', 'I-LOCATION'),
 ('หอ', 'I-LOCATION'),
 (' ', 'B-LOCATION'),
 ('อ', 'B-LOCATION'),
 ('.', 'I-LOCATION'),
 ('ภู', 'I-LOCATION'),
 ('หลวง', 'I-LOCATION'),
 (' ', 'B-LOCATION'),
 ('จ', 'B-LOCATION'),
 ('.', 'I-LOCATION'),
 ('เลย', 'I-LOCATION'),
 (' ', 'B-ZIP'),
 ('4', 'B-ZIP'),
 ('22', 'B-ZIP'),
 ('30', 'B-ZIP')]

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 2e-05
train_batch_size: 16
eval_batch_size: 16
seed: 42
optimizer: Use adamw_torch with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
lr_scheduler_type: linear
num_epochs: 10

Training results

Training Loss	Epoch	Step	Validation Loss	Precision	Recall	F1	Accuracy
No log	1.0	274	0.1790	0.6867	0.7908	0.7351	0.9484
0.2681	2.0	548	0.1331	0.7788	0.8463	0.8111	0.9650
0.2681	3.0	822	0.1135	0.8082	0.8766	0.8410	0.9692
0.0829	4.0	1096	0.1053	0.8300	0.8870	0.8575	0.9717
0.0829	5.0	1370	0.1136	0.8175	0.8868	0.8507	0.9704
0.0512	6.0	1644	0.1135	0.8408	0.8836	0.8616	0.9723
0.0512	7.0	1918	0.1162	0.8429	0.8894	0.8656	0.9725
0.037	8.0	2192	0.1205	0.8475	0.8916	0.8690	0.9730
0.037	9.0	2466	0.1237	0.8490	0.8942	0.8710	0.9732
0.0275	10.0	2740	0.1222	0.8480	0.8934	0.8701	0.9733

Framework versions

Transformers 4.47.1
Pytorch 2.5.1+cu121
Datasets 3.2.0
Tokenizers 0.21.0