|
--- |
|
license: creativeml-openrail-m |
|
language: |
|
- my |
|
tags: |
|
- Myanmar |
|
- Burmese |
|
- GPT2 |
|
- MyanmarGPT |
|
- Nautral Language Processing |
|
widget: |
|
- text: "အီတလီ" |
|
example_title: "Example 1" |
|
- text: "အနုပညာ" |
|
example_title: "Example 2" |
|
- text: "တရုတ်" |
|
example_title: "Example 3" |
|
- text: "ကျောက်ခေတ်" |
|
example_title: "Example 4" |
|
- text: "မြန်မာနိုင်ငံ" |
|
example_title: "Example 5" |
|
--- |
|
|
|
# Myanmar-GPT |
|
|
|
မြန်မာ(ဗမာ)လိုနားလည်သော GPT - Myanmar GPT |
|
|
|
Myanmar GPT is a model trained on a private Myanmar language dataset made by MinSiThu. |
|
The project aims to make the Myanmar language available in the GPT2 Model. |
|
|
|
Fine-tuning the MyanmarGPT model makes it easier to build a custom Myanmar language model than using alternative language models. |
|
|
|
Reports on training the MyanmarGPT model are visualized at [MyanmarGPT Report](https://api.wandb.ai/links/minsithu/wn8yul90). |
|
|
|
There is also 1.42 billion parameters MyanmarGPT-Big model with multilanguage support. |
|
You are find [MyanmarGPT-Big Here](https://huggingface.co/jojo-ai-mst/MyanmarGPT-Big). |
|
|
|
## How to use in your project |
|
|
|
``` |
|
!pip install transformers |
|
``` |
|
|
|
```python |
|
from transformers import pipeline |
|
|
|
generator = pipeline(model="jojo-ai-mst/MyanmarGPT") |
|
outputs = generator("အီတလီ",do_sample=False) |
|
|
|
print(outputs) |
|
# [{'generated_text': 'အီတလီနိုင်ငံသည် ဥရောပတိုက်၏ တောင်ဘက်တွင် မြေထဲပင်လယ်ထဲသို့ ထိုးထွက်နေသော ကျွန်းဆွယ်ကြီးတစ်ခုဖြစ်၍ ပုံသဏ္ဌာန်အားဖြင့် မြင်းစီးဖိနပ်နှင့် တူလေသည်။ မြောက်ဘက်မှ တောင်ဘက်အငူစွန်းအထိ မိုင်ပေါင်း ၇၅ဝ ခန့် ရှည်လျား၍၊ ပျမ်းမျှမိုင် ၁ဝဝ မှ ၁၂ဝ ခန့်ကျယ်သည်။ အီတလီနိုင်ငံ၏ အကျယ်အဝန်းမှာ ဆာဒင်းနီးယားကျွန်း၊ စစ္စလီကျွန်းနှင့် အနီးပတ်ဝန်းကျင်ရှိ ကျွန်းကလေးများ အပါအဝင် ၁၁၆,၃၅၀ စတုရန်းမိုင်ရှိသည်။ '}] |
|
``` |
|
|
|
### alternative ways |
|
|
|
```python |
|
import torch |
|
from transformers import AutoTokenizer, AutoModelForCausalLM |
|
|
|
model = GPT2LMHeadModel.from_pretrained("jojo-ai-mst/MyanmarGPT") |
|
tokenizer = GPT2Tokenizer.from_pretrained("jojo-ai-mst/MyanmarGPT") |
|
|
|
def generate_text(prompt, max_length=300, temperature=0.8, top_k=50): |
|
input_ids = tokenizer.encode(prompt, return_tensors="pt").cuda() # remove .cude() if only cpu |
|
output = model.generate( |
|
input_ids, |
|
max_length=max_length, |
|
temperature=temperature, |
|
top_k=top_k, |
|
pad_token_id=tokenizer.eos_token_id, |
|
do_sample=True |
|
) |
|
for result in output: |
|
generated_text = tokenizer.decode(result, skip_special_tokens=True) |
|
print(generated_text) |
|
|
|
generate_text("အီတလီ ") |
|
``` |
|
|
|
|
|
## Here are the guidelines for using the MyanmarGPT license, |
|
- MyanmarGPT is free to use for everyone, |
|
|
|
- **Must Do** |
|
- any project derived/finetuned from MyanmarGPT, used MyanmarGPT internally, |
|
- or modified MyanmarGPT, related to MyanmarGPT **must mention the citation below** in the corresponding project's page. |
|
- the citation |
|
```latex |
|
@software{MyanmarGPT, |
|
author = {{MinSiThu}}, |
|
title = {MyanmarGPT}, |
|
version={1.1-SweptWood} |
|
url = {https://huggingface.co/jojo-ai-mst/MyanmarGPT}, |
|
urldate = {2023-12-14} |
|
date = {2023-12-14}, |
|
} |
|
``` |
|
|
|
For contact, reach me via [https://www.linkedin.com/in/min-si-thu/](https://www.linkedin.com/in/min-si-thu/) |