---
license: cc-by-nc-sa-4.0
datasets:
- traintogpb/aihub-flores-koen-integrated-sparta-base-300k
language:
- en
- ko
metrics:
- sacrebleu
- xcomet
pipeline_tag: translation
tags:
- translation
- text-generation
- ko2en
- en2ko
---
### Pretrained LM
- [beomi/Llama-3-Open-Ko-8B](https://huggingface.co/beomi/Llama-3-Open-Ko-8B) (MIT License)

### Training Dataset
- [traintogpb/aihub-flores-koen-integrated-prime-base-300k](https://huggingface.co/datasets/traintogpb/aihub-flores-koen-integrated-prime-base-300k)
- Can translate in Enlgish-Korean (bi-directional)

### Prompt
- Template:
  ```python
    prompt = f"Translate this from {src_lang} to {tgt_lang}\n### {src_lang}: {src_text}\n### {tgt_lang}:"

    >>> # src_lang can be 'English', '한국어'
    >>> # tgt_lang can be '한국어', 'English'
  ```
  Mind that there is no "space (`_`)" at the end of the prompt (unpredictable first token will be popped up).

### Training
- Trained with QLoRA
  - PLM: NormalFloat 4-bit
  - Adapter: BrainFloat 16-bit
  - Adapted to all the linear layers (around 2.05%)
- Merge adapters and upscaled in BrainFloat 16-bit precision

### Usage (IMPORTANT)
- Should remove the EOS token (`<|end_of_text|>`, id=128001) at the end of the prompt.
  ```python
    # MODEL
    model_name = 'traintogpb/llama-3-enko-translator-8b-qlora-bf16-upscaled'
    model = AutoModelForCausalLM.from_pretrained(
        model_name,
        max_length=768,
        attn_implementation='flash_attention_2',
        torch_dtype=torch.bfloat16,
    )

    tokenizer = AutoTokenizer.from_pretrained(adapter_name)
    tokenizer.pad_token_id = 128002 # eos_token_id and pad_token_id should be different
    # tokenizer.add_eos_token = False # There is no 'add_eos_token' option in llama3

    text = "Someday, QWER will be the greatest girl band in the world."
    input_prompt = f"Translate this from English to 한국어.\n### English: {text}\n### 한국어:"
    inputs = tokenizer(input_prompt, max_length=768, truncation=True, return_tensors='pt')

    if inputs['input_ids'][0][-1] == tokenizer.eos_token_id:
        inputs['input_ids'] = inputs['input_ids'][0][:-1].unsqueeze(dim=0)
        inputs['attention_mask'] = inputs['attention_mask'][0][:-1].unsqueeze(dim=0)

    outputs = model.generate(**inputs, max_length=768, eos_token_id=tokenizer.eos_token_id)

    input_len = len(inputs['input_ids'].squeeze())
    translation = tokenizer.decode(outputs[0][input_len:], skip_special_tokens=True)
    print(translation)
  ```

### Framework versions

- PEFT 0.8.2