Sponsored by
Vmake Labs Video Enhancer
-Vmake Labs Video Enhancer uses AI to improve video quality, restore details, reduce blur, and upscale videos to 8K.
This model is a fine-tuned version of
NlpHUST/electra-base-vn
on an vlsp 2013 vietnamese word segmentation dataset.
It achieves the following results on the evaluation set:
Loss: 0.0501
Precision: 0.9833
Recall: 0.9838
F1: 0.9835
Accuracy: 0.9911
Model description
More information needed
Intended uses & limitations
You can use this model with Transformers
pipeline
for NER.
from transformers import AutoTokenizer, AutoModelForTokenClassification
from transformers import pipeline
tokenizer = AutoTokenizer.from_pretrained("NlpHUST/vi-word-segmentation")
model = AutoModelForTokenClassification.from_pretrained("NlpHUST/vi-word-segmentation")
nlp = pipeline("token-classification", model=model, tokenizer=tokenizer)
example = "Phát biểu tại phiên thảo luận về tình hình kinh tế xã hội của Quốc hội sáng 28/10 , Bộ trưởng Bộ LĐ-TB&XH Đào Ngọc Dung khái quát , tại phiên khai mạc kỳ họp , lãnh đạo chính phủ đã báo cáo , đề cập tương đối rõ ràng về việc thực hiện các chính sách an sinh xã hội"
ner_results = nlp(example)
example_tok = ""for e in ner_results:
if"##"in e["word"]:
example_tok = example_tok + e["word"].replace("##","")
elif e["entity"] =="I":
example_tok = example_tok + "_" + e["word"]
else:
example_tok = example_tok + " " + e["word"]
print(example_tok)
Phát_biểu tại phiên thảo_luận về tình_hình kinh_tế xã_hội của Quốc_hội sáng 28 / 10 , Bộ_trưởng Bộ LĐ - TB [UNK] XH Đào_Ngọc_Dung khái_quát , tại phiên khai_mạc kỳ họp , lãnh_đạo chính_phủ đã báo_cáo , đề_cập tương_đối rõ_ràng về việc thực_hiện các chính_sách an_sinh xã_hội
Training and evaluation data
More information needed
Training procedure
Training hyperparameters
The following hyperparameters were used during training:
learning_rate: 5e-05
train_batch_size: 8
eval_batch_size: 4
seed: 42
gradient_accumulation_steps: 2
total_train_batch_size: 16
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: linear
num_epochs: 5.0
Training results
Training Loss
Epoch
Step
Validation Loss
Precision
Recall
F1
Accuracy
0.0168
1.0
4712
0.0284
0.9813
0.9825
0.9819
0.9904
0.0107
2.0
9424
0.0350
0.9789
0.9814
0.9802
0.9895
0.005
3.0
14136
0.0364
0.9826
0.9843
0.9835
0.9909
0.0033
4.0
18848
0.0434
0.9830
0.9831
0.9830
0.9908
0.0017
5.0
23560
0.0501
0.9833
0.9838
0.9835
0.9911
Framework versions
Transformers 4.22.2
Pytorch 1.12.1+cu113
Datasets 2.4.0
Tokenizers 0.12.1
Runs of NlpHUST vi-word-segmentation on huggingface.co
1.6K
Total runs
0
24-hour runs
0
3-day runs
0
7-day runs
0
30-day runs
More Information About vi-word-segmentation huggingface.co Model
vi-word-segmentation huggingface.co
vi-word-segmentation huggingface.co is an AI model on huggingface.co that provides vi-word-segmentation's model effect (), which can be used instantly with this NlpHUST vi-word-segmentation model. huggingface.co supports a free trial of the vi-word-segmentation model, and also provides paid use of the vi-word-segmentation. Support call vi-word-segmentation model through api, including Node.js, Python, http.
vi-word-segmentation huggingface.co is an online trial and call api platform, which integrates vi-word-segmentation's modeling effects, including api services, and provides a free online trial of vi-word-segmentation, you can try vi-word-segmentation online for free by clicking the link below.
NlpHUST vi-word-segmentation online free url in huggingface.co:
vi-word-segmentation is an open source model from GitHub that offers a free installation service, and any user can find vi-word-segmentation on GitHub to install. At the same time, huggingface.co provides the effect of vi-word-segmentation install, users can directly use vi-word-segmentation installed effect in huggingface.co for debugging and trial. It also supports api for free installation.
vi-word-segmentation install url in huggingface.co: