Khmer Part of Speech Tagging with XLM RoBERTa

This model is a fine-tuned version of xlm-roberta-base on the khPOS dataset. It achieves the following results on the evaluation set:

  • Loss: 0.1063
  • Precision: 0.9512
  • Recall: 0.9526
  • F1: 0.9519
  • Accuracy: 0.9735

Model description

The original paper achieved 98.15% accuracy while this model achieved only 97.35% which is close. However, this is a multilingual model so it has more tokens than the original paper.

Intended uses & limitations

This model can be used to extract useful information from Khmer text.

Training and evaluation data

train: 90% / test: 10%

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

  • learning_rate: 2e-05
  • train_batch_size: 24
  • eval_batch_size: 16
  • 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 Precision Recall F1 Accuracy
No log 1.0 450 0.1347 0.9314 0.9333 0.9324 0.9603
0.4834 2.0 900 0.1183 0.9407 0.9377 0.9392 0.9653
0.1323 3.0 1350 0.1026 0.9484 0.9482 0.9483 0.9699
0.095 4.0 1800 0.0986 0.9502 0.9490 0.9496 0.9712
0.0774 5.0 2250 0.0978 0.9494 0.9491 0.9493 0.9712
0.0616 6.0 2700 0.0991 0.9493 0.9507 0.9500 0.9715
0.0494 7.0 3150 0.0989 0.9529 0.9540 0.9534 0.9731
0.0414 8.0 3600 0.1037 0.9499 0.9501 0.9500 0.9722
0.0339 9.0 4050 0.1056 0.9516 0.9517 0.9516 0.9734
0.029 10.0 4500 0.1063 0.9512 0.9526 0.9519 0.9735

Framework versions

  • Transformers 4.30.2
  • Pytorch 2.0.1+cu118
  • Datasets 2.13.1
  • Tokenizers 0.13.3
Downloads last month
20
Safetensors
Model size
277M params
Tensor type
I64
·
F32
·
Inference Examples
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social visibility and check back later, or deploy to Inference Endpoints (dedicated) instead.

Dataset used to train seanghay/khmer-pos-roberta

Evaluation results