traintogpb/llama-3-enko-translator-8b-qlora-bf16-upscaled

Pretrained LM

• beomi/Llama-3-Open-Ko-8B (MIT License)

Training Dataset

• traintogpb/aihub-flores-koen-integrated-prime-base-300k
• Can translate in Enlgish-Korean (bi-directional)

Prompt

• Template:

    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.

    # 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