Hugging Face's logo

simbolo-ai
/
Myanmarsar-GPT

Text Generation
Transformers
Safetensors
Burmese
gpt2
burmese
pre-trained
text-generation-inference
Inference Endpoints
Model card Files Community
Myanmarsar-GPT / README.md
Simbolo's picture
Simbolo
Update README.md
aa8672c verified
preview code
|
raw
history blame
3.34 kB
metadata
license: mit
language:
  - my
pipeline_tag: text-generation
metrics:
  - code_eval
library_name: transformers
tags:
  - burmese
  - gpt2
  - pre-trained

widget:

  • text: 'ပညာရေး

The Simbolo's Myanmarsar-GPT (it is not a chatbot but a text generation model which can be used to develop chatbot) is pre-trained on a dataset of 20,000 Burmese data and pre-trained using the GPT-2 architecture of MGPT Model. Its purpose is to serve as a foundational pre-trained model for the Burmese language, facilitating fine-tuning for specific applications of different tasks such as creative writing, chatbot, machine translation etc.

image/jpeg

How to use 

from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("Simbolo-Servicio/Myanmarsar-GPT")
model = AutoModelForCausalLM.from_pretrained("Simbolo-Servicio/Myanmarsar-GPT")

input_text = "ပညာရေး"
input_ids = tokenizer.encode(input_text, return_tensors='pt')
output = model.generate(input_ids, max_length=50)
print(tokenizer.decode(output[0], skip_special_tokens=True))

Data

We use 20,000 Burmese sentences and most are from our open-source data which contains 100,000 sentences sourced from Wikipedia.

Contributors

Main Contributor: Sa Phyo Thu Htet Wikipedia Data Crawling: Kaung Kaung Ko Ko, Phuu Pwint Thinzar Kyaing Releasing the Model: Eithandaraung, Ye Yint Htut, Thet Chit Su, Naing Phyo Aung, Nyan Linn Phyo Zaw, Lynn Thu Kha

Acknowledgment

We extend our gratitude to the creators of the mGPT-XL models for their invaluable contribution to this project. We want to thank everyone who has worked on the related works, especially Minsithu and Dr. Wai Yan Nyein Naingwho initiated the work of gpt-2 model. And We would like to thank Simbolo:Servico which is a branch of Simbolo under the company of Intello Tech for providing financial support.

Limitations and Bias

We have yet to investigate the potential bias inherent in this model thoroughly. Regarding transparency, it's important to note that the model is primarily trained on data from the Unicode Burmese(Myanmar) language.

References

  1. Jiang, Shengyi & Huang, Xiuwen & Cai, Xiaonan & Lin, Nankai. (2021). Pre-trained Models and Evaluation Data for the Myanmar Language. 10.1007/978-3-030-92310-5_52.
  2. Lin, N., Fu, Y., Chen, C., Yang, Z., & Jiang, S. (2021). LaoPLM: Pre-trained Language Models for Lao. ArXiv. /abs/2110.05896
  3. MinSithu, MyanmarGPT, https://huggingface.co/jojo-ai-mst/MyanmarGPT, 1.1-SweptWood
  4. Wai Yan Nyein Naing, WYNN747/Burmese-GPT, https://huggingface.co/WYNN747/Burmese-GPT
  5. Sai Htaung Kham, saihtaungkham/BurmeseRoBERTaCLM
  6. Shliazhko, O., Fenogenova, A., Tikhonova, M., Mikhailov, V., Kozlova, A., & Shavrina, T. (2022). MGPT: Few-Shot Learners Go Multilingual. ArXiv. /abs/2204.07580

How to Cite this work:

Sa Phyo Thu Htet, Simbolo (2023). Myanmarsar-GPT, https://huggingface.co/Simbolo-Servicio/Myanmarsar-GPT/