Hugging Face's logo

youjunhyeok
/
llama3-8B-dpo-sft-v1

Text Generation
Transformers
Safetensors
llama
llama-factory
conversational
text-generation-inference
Inference Endpoints
Model card Files Community
llama3-8B-dpo-sft-v1
File size: 3,230 Bytes 
2676805
 
93ab893
 
 
2676805
d97843a
 
 
2676805
93ab893
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
 
---
library_name: transformers
tags:
- llama-factory
license: apache-2.0
---
## Info
SFT > DPO 순서가 아닌 DPO > SFT 순서로 학습시킨 모델입니다.
SFT > DPO는 [여기](https://huggingface.co/youjunhyeok/llama3-8b-ko-sft-dpo-v1)에서 확인해 주세요.

## Model
- base model: [meta-llama/Meta-Llama-3-8B-Instruct](https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct)

## Dataset
- [youjunhyeok/llama3_train](https://huggingface.co/datasets/youjunhyeok/llama3_train)
- [youjunhyeok/ko-orca-pair-and-ultrafeedback-dpo](https://huggingface.co/datasets/youjunhyeok/ko-orca-pair-and-ultrafeedback-dpo)

## Load Model

Use the following Python code to load the model:

```python3
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "youjunhyeok/llama3-8B-dpo-sft-v1"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
```

## Chat

```python3
def chat(message):
    messages = [
        {"role": "system", "content": "당신은 친절하고 도움이 되는 챗봇입니다."},
        {"role": "user", "content": message},
    ]

    input_ids = tokenizer.apply_chat_template(
        messages,
        add_generation_prompt=True,
        return_tensors="pt"
    ).to(model.device)

    terminators = [
        tokenizer.eos_token_id,
        tokenizer.convert_tokens_to_ids("<|eot_id|>")
    ]

    outputs = model.generate(
        input_ids,
        max_new_tokens=512,
        eos_token_id=terminators,
        do_sample=False,
        temperature=0.5,
        top_p=0.8,
    )
    response = outputs[0][input_ids.shape[-1]:]
    print(tokenizer.decode(response, skip_special_tokens=True))

chat('한산도 대첩에 대해 아는 대로 얘기해봐')
```

## Output

```
한산도 대첩은 조선시대에 일어난 전투로, 이순신 장군이 이끄는 조선군이 일본군을 물리친 전투입니다.
```

## BenchMark (KOR)

```
# alias
A = youjunhyeok/llama3-8B-dpo-sft-v1
B = DavidAhn/Llama-3-8B-slerp-262k
C = meta-llama/Meta-Llama-3-8B
D = chihoonlee10/T3Q-ko-solar-dpo-v7.0 (24.05.24 ko 리더보드 1등)
```

| Benchmark (macro_f1)      |   A  |   B  |   C  |   D  |
|---------------------------|:----:|:----:|:----:|:----:|
| kobest_boolq (0-shot)     | 84.7 | 33.5 | 38.2 | 34.1 |
| kobest_boolq (5-shot)     | 86.1 | 68.8 | 83.8 | 93.1 |
| kobest_copa (0-shot)      | 60.6 | 58.5 | 63.1 | 81.0 |
| kobest_copa (5-shot)      | 67.2 | 61.7 | 69.1 | 91.0 |
| kobest_hellaswag (0-shot) | 40.0 | 43.2 | 42.1 | 55.1 |
| kobest_hellaswag (5-shot) | 42.4 | 45.3 | 44.2 | 55.2 |
| kobest_sentineg (0-shot)  | 52.1 | 34.8 | 51.5 | 82.7 |
| kobest_sentineg (5-shot)  | 89.4 | 85.8 | 94.7 | 91.4 |

## BenchMark (ENG)


|                                               |        openbookqa |        hellaswag |       boolq |       arc_easy | arc_challenge | 
|:----------------------------------------------|---------:|---------:|---------:|---------:|---------:|
| youjunhyeok/llama3-8B-dpo-sft-v1              | 0.320 | 0.547 | 0.529 | 0.748 | 0.446 |
| meta-llama/Meta-Llama-3-8B-Instruct           | 0.338 | 0.576 | 0.831 | 0.815 | 0.529 |