File size: 3,384 Bytes
2637e65 |
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 |
---
language:
- en
license: llama3.1
tags:
- llama-3.1
- ncsoft
- varco
base_model:
- meta-llama/Meta-Llama-3.1-8B-Instruct
---
## Llama-3.1-Varco-8B-Instruct
### About the Model
**Llama-3.1-Varco-8B-Instruct** is a *generative model* based on Meta-Llama-3.1-8B, specifically designed to excel in Korean through additional training. The model uses continual pre-training with both Korean and English datasets to enhance its understanding and generation capabilites in Korean, while also maintaining its proficiency in English. It performs supervised fine-tuning (SFT) and direct preference optimization (DPO) in Korean to align with human preferences.
- **Developed by:** NC Research, Language Model Team
- **Languages (NLP):** Korean, English
- **License:** LLAMA 3.1 COMMUNITY LICENSE AGREEMENT
- **Base model:** [meta-llama/Meta-Llama-3.1-8B](https://huggingface.co/meta-llama/Meta-Llama-3.1-8B)
## Uses
### Direct Use
We recommend to use transformers v4.43.0 or later, as advised for Llama-3.1.
```python
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
model = AutoModelForCausalLM.from_pretrained(
"NCSOFT/Llama-3.1-Varco-8B-Instruct",
torch_dtype=torch.bfloat16,
device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("NCSOFT/Llama-3.1-Varco-8B-Instruct")
messages = [
{"role": "system", "content": "You are a helpful assistant Varco. Respond accurately and diligently according to the user's instructions."},
{"role": "user", "content": "์๋
ํ์ธ์."}
]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt").to(model.device)
eos_token_id = [
tokenizer.eos_token_id,
tokenizer.convert_tokens_to_ids("<|eot_id|>")
]
outputs = model.generate(
inputs,
eos_token_id=eos_token_id,
max_length=8192
)
print(tokenizer.decode(outputs[0]))
```
## Evaluation
### LogicKor
We used the [LogicKor](https://github.com/instructkr/LogicKor) code to measure performance. For the judge model, we used the officially recommended gpt-4-1106-preview. The score includes only the 0-shot evaluation provided in the default.
| Model | Math | Reasoning | Writing | Coding | Understanding | Grammer | Single turn | Multi turn | Overall |
|--------------|--------|-------------|-----------|----------|-----------------|-----------|---------------|--------------|-----------|
| [Llama-3.1-Varco-8B-Instruct](https://huggingface.co/NCSOFT/Llama-3.1-Varco-8B-Instruct)| 6.71 / 8.57 | 8.86 / 8.29 | 9.86 / 9.71 | 8.86 / 9.29 | 9.29 / 10.0 | 8.57 / 7.86 | 8.69 | 8.95 | 8.82 |
| [EXAONE-3.0-7.8B-Instruct](https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct)| 6.86 / 7.71 | 8.57 / 6.71 | 10.0 / 9.29 | 9.43 / 10.0 | 10.0 / 10.0 | 9.57 / 5.14 | 9.07 | 8.14 | 8.61 |
| [Meta-Llama-3.1-8B-Instruct](https://huggingface.co/meta-llama/Meta-Llama-3.1-8B-Instruct)| 4.29 / 4.86 | 6.43 / 6.57 | 6.71 / 5.14 | 6.57 / 6.00 | 4.29 / 4.14 | 6.00 / 4.00 | 5.71 | 5.12 | 5.42 |
| [Gemma-2-9B-Instruct](https://huggingface.co/google/gemma-2-9b-it)| 6.14 / 5.86 | 9.29 / 9.0 | 9.29 / 8.57 | 9.29 / 9.14 | 8.43 / 8.43 | 7.86 / 4.43 | 8.38 | 7.57 | 7.98
| [Qwen2-7B-Instruct](https://huggingface.co/Qwen/Qwen2-7B-Instruct)| 5.57 / 4.86 | 7.71 / 6.43 | 7.43 / 7.00 | 7.43 / 8.00 | 7.86 / 8.71 | 6.29 / 3.29 | 7.05 | 6.38 | 6.71 |
|