File size: 2,865 Bytes
8347832
 
 
8fad968
 
 
 
 
 
 
8347832
945d9c8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1faf122
945d9c8
 
1163aed
945d9c8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
bfd119d
945d9c8
 
 
 
 
 
 
 
9101027
 
 
 
945d9c8
8347832
 
 
1163aed
1faf122
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
---
library_name: peft
base_model: beomi/open-llama-2-ko-7b
license: cc-by-sa-4.0
datasets:
- traintogpb/aihub-flores-koen-integrated-sparta-mini-300k
language:
- en
- ko
pipeline_tag: translation
---
### 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-sparta-mini-300k](https://huggingface.co/datasets/traintogpb/aihub-flores-koen-integrated-sparta-mini-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 a "space (`_`)" at the end of the prompt (unpredictable first token will be popped up).
  But if you use vLLM, it's okay to remove the final space(`_`).

### 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 at the end of the prompt.
  ```python
    # MODEL
    model_name = 'beomi/Llama-3-Open-Ko-8B'
    adapter_name = 'traintogpb/llama-3-enko-translator-8b-qlora-adapter'
    bnb_config = BitsAndBytesConfig(
        load_in_4bit=True,
        bnb_4bit_quant_type='nf4',
        bnb_4bit_compute_dtype=torch.bfloat16,
        bnb_4bit_use_double_quant=True
    )
    model = AutoModelForCausalLM.from_pretrained(
        model_name,
        max_length=768,
        quantization_config=bnb_config,
        attn_implementation='flash_attention_2',
        torch_dtype=torch.bfloat16,
    )
    model = PeftModel.from_pretrained(
        model,
        adapter_path=adapter_name,
        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

    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