README.md

---

license: cc-by-nc-sa-4.0
datasets:
- nlpai-lab/kullm-v2
language:
- ko
pipeline_tag: text-generation
---


교육용으로 학습 한 간단한 instruction fine-tuning 모델 (updated 2023/08/06)

- Pretrained model: skt/kogpt2-base-v2 (https://github.com/SKT-AI/KoGPT2)
- Training data: kullm-v2(https://huggingface.co/datasets/nlpai-lab/kullm-v2)

```python

from transformers import AutoModelForCausalLM

from transformers import PreTrainedTokenizerFast



tokenizer = PreTrainedTokenizerFast.from_pretrained("hyunjae/skt-kogpt2-kullm-v2",

                                                    bos_token='</s>', eos_token='</s>', unk_token='<unk>',

                                                    pad_token='<pad>', mask_token='<mask>', padding_side="right", model_max_length=512)

model = AutoModelForCausalLM.from_pretrained('hyunjae/skt-kogpt2-kullm-v2').to('cuda')



PROMPT= "### system:사용자의 질문에 맞는 적절한 응답을 생성하세요.\n### 사용자:{instruction}\n### 응답:"

text = PROMPT.format_map({'instruction':"안녕? 너가 할 수 있는게 뭐야?"})

input_ids = tokenizer.encode(text, return_tensors='pt').to(model.device)



gen_ids = model.generate(input_ids,

                        repetition_penalty=2.0,

                        pad_token_id=tokenizer.pad_token_id,

                        eos_token_id=tokenizer.eos_token_id,

                        bos_token_id=tokenizer.bos_token_id,

                        num_beams=4,

                        no_repeat_ngram_size=4,

                        max_new_tokens=128,

                        do_sample=True,

                        top_k=50)





generated = tokenizer.decode(gen_ids[0])

print(generated)



```