5CD-AI/viso-twhin-bert-large

Overview

We reduce TwHIN-BERT's vocabulary size to 20k on the UIT dataset and continue pretraining for 10 epochs.

Here are the results on 4 downstream tasks on Vietnamese social media texts, including Emotion Recognition(UIT-VSMEC), Hate Speech Detection(UIT-HSD), Spam Reviews Detection(ViSpamReviews), Hate Speech Spans Detection(ViHOS):

Model	Avg	Emotion Recognition			Hate Speech Detection			Spam Reviews Detection			Hate Speech Spans Detection
		Acc	WF1	MF1	Acc	WF1	MF1	Acc	WF1	MF1	Acc	WF1	MF1
viBERT	78.16	61.91	61.98	59.7	85.34	85.01	62.07	89.93	89.79	76.8	90.42	90.45	84.55
vELECTRA	79.23	64.79	64.71	61.95	86.96	86.37	63.95	89.83	89.68	76.23	90.59	90.58	85.12
PhoBERT-Base	79.3	63.49	63.36	61.41	87.12	86.81	65.01	89.83	89.75	76.18	91.32	91.38	85.92
PhoBERT-Large	79.82	64.71	64.66	62.55	87.32	86.98	65.14	90.12	90.03	76.88	91.44	91.46	86.56
ViSoBERT	81.58	68.1	68.37	65.88	88.51	88.31	68.77	90.99	90.92	79.06	91.62	91.57	86.8
visobert-14gb-corpus	82.2	68.69	68.75	66.03	88.79	88.6	69.57	91.02	90.88	77.13	93.69	93.63	89.66
viso-twhin-bert-large	83.87	73.45	73.14	70.99	88.86	88.8	70.81	91.6	91.47	79.07	94.08	93.96	90.22

Usage (HuggingFace Transformers)

Install transformers package:

pip install transformers

Then you can use this model for fill-mask task like this:

from transformers import pipeline

model_path = "5CD-AI/viso-twhin-bert-large"
mask_filler = pipeline("fill-mask", model_path)

mask_filler("đúng nhận sai <mask>", top_k=10)

Fine-tune Configuration

We fine-tune 5CD-AI/viso-twhin-bert-large on 4 downstream tasks with transformer library with the following configuration:

• train_batch_size: 16
• seed: 42
• gradient_accumulation_steps: 4
• weight_decay: 0.01
• optimizer: AdamW with betas=(0.9, 0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• training_epochs: 30
• model_max_length: 128
• metric_for_best_model: wf1
• strategy: epoch

And different additional configurations for each task:

Emotion Recognition	Hate Speech Detection	Spam Reviews Detection	Hate Speech Spans Detection
- learning_rate: 1e-5	- learning_rate: 5e-6	- learning_rate: 1e-5	- learning_rate: 5e-6