metadata

license: mit
base_model: microsoft/mdeberta-v3-base
tags:
  - generated_from_trainer
datasets:
  - universal_dependencies
metrics:
  - accuracy
  - precision
  - recall
model-index:
  - name: mdeberta-v3-ud-thai-pud-upos
    results:
      - task:
          name: Token Classification
          type: token-classification
        dataset:
          name: universal_dependencies
          type: universal_dependencies
          config: th_pud
          split: test
          args: th_pud
        metrics:
          - name: Accuracy
            type: accuracy
            value: 0.9934846474601972
widget:
  - text: >-
      นักวิจัยกล่าวว่าการวิเคราะห์ดีเอ็นเอของเนื้องอกอาจช่วยอธิบายถึงสาเหตุที่แท้จริงของมะเร็งชนิดอื่นๆ
      ได้
    example_title: test_example_1
  - text: คือผมไม่ได้ชอบกดดันพวกคุณหรอกนะ แต่ชะตากรรมของสาธารณรัฐอยู่ในกำมือคุณ
    example_title: test_example_2
language:
  - th
library_name: transformers

mdeberta-v3-ud-thai-pud-upos

This model is a fine-tuned version of microsoft/mdeberta-v3-base on the universal_dependencies dataset. It achieves the following results on the evaluation set:

Loss: 0.0303
Macro avg precision: 0.9235
Macro avg recall: 0.9228
Macro avg f1: 0.9231
Weighted avg precision: 0.9935
Weighted avg recall: 0.9935
Weighted avg f1: 0.9935
Accuracy: 0.9935

Model description

This model is train on thai UD Thai PUD corpus with Universal Part-of-speech (UPOS) tag to help with pos tagging in Thai language.

Example

from transformers import AutoModelForTokenClassification, AutoTokenizer, TokenClassificationPipeline

model = AutoModelForTokenClassification.from_pretrained("Pavarissy/mdeberta-v3-ud-thai-pud-upos")
tokenizer = AutoTokenizer.from_pretrained("Pavarissy/mdeberta-v3-ud-thai-pud-upos")

pipeline = TokenClassificationPipeline(model=model, tokenizer=tokenizer, grouped_entities=True)
outputs = pipeline("ประเทศไทย อยู่ใน ทวีป เอเชีย")
print(outputs)
# [{'entity_group': 'PROPN', 'score': 0.9946701, 'word': 'ประเทศไทย', 'start': 0, 'end': 9}, {'entity_group': 'VERB', 'score': 0.85809743, 'word': 'อยู่ใน', 'start': 9, 'end': 16}, {'entity_group': 'NOUN', 'score': 0.99632, 'word': 'ทวีป', 'start': 16, 'end': 21}, {'entity_group': 'PROPN', 'score': 0.9961184, 'word': 'เอเชีย', 'start': 21, 'end': 28}]

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 2e-05
train_batch_size: 8
eval_batch_size: 8
seed: 42
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: linear
num_epochs: 10

Training results

Training Loss	Epoch	Step	Validation Loss	Macro avg precision	Macro avg recall	Macro avg f1	Weighted avg precision	Weighted avg recall	Weighted avg f1	Accuracy
No log	1.0	125	0.3898	0.8417	0.7849	0.8078	0.9119	0.9112	0.9101	0.9112
No log	2.0	250	0.1768	0.8765	0.8683	0.8720	0.9561	0.9560	0.9559	0.9560
No log	3.0	375	0.1217	0.8972	0.8892	0.8929	0.9701	0.9701	0.9699	0.9701
0.4709	4.0	500	0.0841	0.9057	0.9064	0.9059	0.9802	0.9800	0.9800	0.9800
0.4709	5.0	625	0.0649	0.9128	0.9133	0.9130	0.9854	0.9853	0.9853	0.9853
0.4709	6.0	750	0.0513	0.9147	0.9170	0.9158	0.9878	0.9877	0.9877	0.9877
0.4709	7.0	875	0.0423	0.9199	0.9180	0.9189	0.9900	0.9900	0.9900	0.9900
0.0857	8.0	1000	0.0350	0.9226	0.9207	0.9216	0.9921	0.9921	0.9921	0.9921
0.0857	9.0	1125	0.0318	0.9237	0.9219	0.9228	0.9932	0.9932	0.9932	0.9932
0.0857	10.0	1250	0.0303	0.9235	0.9228	0.9231	0.9935	0.9935	0.9935	0.9935

Framework versions

Transformers 4.34.1
Pytorch 2.1.0+cu118
Datasets 2.14.6
Tokenizers 0.14.1