---
metrics:
- accuracy
model-index:
- name: gpt2-lihkg
  results:
  - task:
      name: Causal Language Modeling
      type: text-generation
    dataset:
      name: lihkg_data
      type: lihkg_data
    metrics:
    - name: Perplexity 
      type: Perplexity 
      value: 30.93
license: openrail
---



# gpt2-shikoto

**Please be aware that the training data might contain inappropriate content. This model is intended for research purposes only.**



The base model can be found [here](https://huggingface.co/jed351/gpt2-base-zh-hk), which was obtained by 
patching a [GPT2 Chinese model](https://huggingface.co/ckiplab/gpt2-base-chinese) and its tokenizer with Cantonese characters.
Refer to the base model for info on the patching process.


The training data was obtained from scrapping an online forum in Hong Kong called LIHKG. 
The tool can be found [here](https://github.com/ayaka14732/lihkg-scraper). 
Please also check out the [Bart model](https://huggingface.co/Ayaka/bart-base-cantonese) created by her.



### Limitations
The model was trained on ~10GB of data scrapped from LIHKG. 
It might contain violent and rude languages so as the text generated by the model. 
Please do not use it for anything other than research or entertainment.


The comments on LIHKG also tend to be very short. 
Thus the model cannot generate anything more than a line. In a lot of occasions might not even generate new tokens.



### Training procedure

Please refer to the [script](https://github.com/huggingface/transformers/tree/main/examples/pytorch/language-modeling) 
provided by Huggingface. 


The model was trained for 400,000 steps with batch size 5 (~2epoches) on 2 NVIDIA Quadro RTX6000 for around 40 hours at the Research Computing Services of Imperial College London. 


### How to use it?
```
from transformers import AutoTokenizer
from transformers import TextGenerationPipeline, AutoModelForCausalLM
tokenizer = AutoTokenizer.from_pretrained("jed351/gpt2_base_zh-hk-lihkg")
model = AutoModelForCausalLM.from_pretrained("jed351/gpt2_base_zh-hk-lihkg")
# try messing around with the parameters
generator = TextGenerationPipeline(model, tokenizer, 
                                   max_new_tokens=200, 
                                   no_repeat_ngram_size=3) #, device=0) #if you have a GPU
input_string = "your input" 
output = generator(input_string)
string = output[0]['generated_text'].replace(' ', '')
print(string)
```


### Framework versions

- Transformers 4.26.0.dev0
- Pytorch 1.13.1
- Datasets 2.8.0
- Tokenizers 0.13.2