metadata

license: apache-2.0
datasets:
  - squarelike/sharegpt_deepl_ko_translation
language:
  - en
  - ko
pipeline_tag: translation

Gugugo-koen-7B-V1.1-GPTQ

Detail repo: https://github.com/jwj7140/Gugugo

This is GPTQ model from squarelike/Gugugo-koen-7B-V1.1

Base Model: Llama-2-ko-7b

Training Dataset: sharegpt_deepl_ko_translation.

I trained with 1x A6000 GPUs for 90 hours.

Prompt Template

KO->EN

### 한국어: {sentence}</끝>
### 영어:

EN->KO

### 영어: {sentence}</끝>
### 한국어:

Implementation Code

from transformers import AutoModelForCausalLM, AutoTokenizer, StoppingCriteria, StoppingCriteriaList
import torch
repo = "squarelike/Gugugo-koen-7B-V1.1-GPTQ"
model = AutoModelForCausalLM.from_pretrained(
        repo,
        device_map='auto'
)
tokenizer = AutoTokenizer.from_pretrained(repo)

model.eval()
model.config.use_cache = True

class StoppingCriteriaSub(StoppingCriteria):
    def __init__(self, stops = [], encounters=1):
        super().__init__()
        self.stops = [stop for stop in stops]

    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor):
        for stop in self.stops:
            if torch.all((stop == input_ids[0][-len(stop):])).item():
                return True

        return False

stop_words_ids = torch.tensor([[829, 45107, 29958], [1533, 45107, 29958], [829, 45107, 29958], [21106, 45107, 29958]]).to("cuda")
stopping_criteria = StoppingCriteriaList([StoppingCriteriaSub(stops=stop_words_ids)])

def gen(lan="en", x=""):
    if (lan == "ko"):
        prompt = f"### 한국어: {x}</끝>\n### 영어:"
    else:
        prompt = f"### 영어: {x}</끝>\n### 한국어:"
    gened = model.generate(
        **tokenizer(
            prompt,
            return_tensors='pt',
            return_token_type_ids=False
        ).to("cuda"),
        max_new_tokens=2000,
        temperature=0.3,
        # no_repeat_ngram_size=5,
        num_beams=5,
        stopping_criteria=stopping_criteria
    )
    return tokenizer.decode(gened[0][1:]).replace(prompt+" ", "").replace("</끝>", "")


print(gen(lan="en", x="Hello, world!"))