llama-2-enko-translator-7b-qlora-adapter

Pretrained LM

• beomi/open-llama-2-ko-7b (MIT License)

Training Dataset

• traintogpb/aihub-flores-koen-integrated-sparta-30k
• 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'
  

• Issue: The tokenizer of the model tokenizes the prompt below in different way with the prompt above. Make sure to use the prompt proposed above.

    prompt = f"""Translate this from {src_lang} to {tgt_lang}
    ### {src_lang}: {src_text}
    ### {tgt_lang}:"""
  
    >>> # DO NOT USE this prompt
  

  And mind that there is no "space (_)" at the end of the prompt.

Training

• Trained with QLoRA
  • PLM: NormalFloat 4-bit
  • Adapter: BrainFloat 16-bit
  • Adapted to all the linear layers (around 2.2%)

Usage (IMPORTANT)

• Should remove the EOS token (<|endoftext|>, id=46332) at the end of the prompt.

    # MODEL
    plm_name = 'beomi/open-llama-2-ko-7b'
    adapter_name = 'traintogpb/llama-2-enko-translator-7b-qlora-adapter'
    model = LlamaForCausalLM.from_pretrained(
        plm_name, 
        max_length=768,
        quantization_config=bnb_config, # Use the QLoRA config above
        attn_implementation='flash_attention_2',
        torch_dtype=torch.bfloat16
    )
    model = PeftModel.from_pretrained(
        model, 
        adapter_name, 
        torch_dtype=torch.bfloat16
    )
  
    # TOKENIZER
    tokenizer = LlamaTokenizer.from_pretrained(plm_name)
    tokenizer.pad_token = "</s>"
    tokenizer.pad_token_id = 2
    tokenizer.eos_token = "<|endoftext|>" # Must be differentiated from the PAD token
    tokenizer.eos_token_id = 46332
    tokenizer.add_eos_token = True
    tokenizer.model_max_length = 768
  
    # INFERENCE
    text = "NMIXX is the world-best female idol group, who came back with the new song 'DASH'." 
    prompt = f"Translate this from {src_lang} to {tgt_lang}\n### {src_lang}: {src_text}\n### {tgt_lang}:"
  
    inputs = tokenizer(prompt, return_tensors="pt", max_length=max_length, truncation=True)
    # REMOVE EOS TOKEN IN THE PROMPT
    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=max_length, eos_token_id=46332)
  
    input_len = len(inputs['input_ids'].squeeze())
            
    translated_text = tokenizer.decode(outputs[0][input_len:], skip_special_tokens=True)
    print(translated_text)