|
--- |
|
language: |
|
- en |
|
- ko |
|
license: cc-by-nc-4.0 |
|
tags: |
|
- dnotitia |
|
- nlp |
|
- llm |
|
- slm |
|
- conversation |
|
- chat |
|
base_model: |
|
- meta-llama/Meta-Llama-3.1-8B |
|
library_name: transformers |
|
pipeline_tag: text-generation |
|
--- |
|
|
|
# DNA 1.0 8B Instruct |
|
|
|
<p align="center"> |
|
<img src="assets/dna-logo.png" width="400" style="margin: 40px auto;"> |
|
</p> |
|
|
|
**DNA 1.0 8B Instruct** is a <u>state-of-the-art (**SOTA**)</u> bilingual language model based on Llama architecture, specifically optimized for Korean language understanding and generation, while also maintaining strong English capabilities. The model was developed through a sophisticated process involving model merging via spherical linear interpolation (**SLERP**) with Llama 3.1 8B Instruct, and underwent knowledge distillation (**KD**) using Llama 3.1 405B as the teacher model. It was extensively trained through continual pre-training (**CPT**) with a high-quality Korean dataset. The training pipeline was completed with supervised fine-tuning (**SFT**) and direct preference optimization (**DPO**) to align with human preferences and enhance instruction-following abilities. |
|
|
|
<p align="center"> |
|
<img src="assets/training-procedure.png" width="600" style="margin: 40px auto;"> |
|
</p> |
|
|
|
DNA 1.0 8B Instruct was fine-tuned on approximately 7B tokens of carefully curated data and has undergone extensive instruction tuning to enhance its ability to follow complex instructions and engage in natural conversations. |
|
|
|
For more details, please refer to our [Technical Report](https://arxiv.org/abs/2501.10648). |
|
|
|
- **Developed by:** Dnotitia Inc. |
|
- **Supported Languages:** Korean, English |
|
- **Model Release Date:** Dec 10, 2024 |
|
- **Vocab Size:** 128,256 |
|
- **Context Length:** 131,072 tokens (128k) |
|
- **License:** CC BY-NC 4.0 |
|
|
|
<div style="padding: 2px 8px; background-color: hsl(240, 100%, 50%, 0.1); border-radius: 5px"> |
|
<p><strong>NOTICE (Korean):</strong></p> |
|
<p>λ³Έ λͺ¨λΈμ μμ
μ λͺ©μ μΌλ‘ νμ©νμ€ μ μμ΅λλ€. μμ
μ μ΄μ©μ μνμλ κ²½μ°, <a href="https://www.dnotitia.com/contact/post-form">Contact us</a>λ₯Ό ν΅ν΄ λ¬Έμν΄ μ£ΌμκΈ° λ°λλλ€. κ°λ¨ν νμ μ μ°¨λ₯Ό κ±°μ³ μμ
μ νμ©μ μΉμΈν΄ λ리λλ‘ νκ² μ΅λλ€.</p> |
|
<p>Try DNA-powered Mnemos Assistant! <a href="https://request-demo.dnotitia.ai/">Beta Open β</a></p> |
|
</div> |
|
|
|
## Evaluation |
|
|
|
We evaluated DNA 1.0 8B Instruct against other prominent language models of similar size across various benchmarks, including Korean-specific tasks and general language understanding metrics. |
|
|
|
| Language | Benchmark | **dnotitia/Llama-DNA-1.0-8B-Instruct** | LGAI-EXAONE/EXAONE-3.5-7.8B-Instruct | LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct | yanolja/EEVE-Korean-Instruct-10.8B-v1.0 | Qwen/Qwen2.5-7B-Instruct | meta-llama/Llama-3.1-8B-Instruct | mistralai/Mistral-7B-Instruct-v0.3 | NCSOFT/Llama-VARCO-8B-Instruct | upstage/SOLAR-10.7B-Instruct-v1.0 | |
|
|----------|------------|----------------------------------------|--------------------------------------|--------------------------------------|-----------------------------------------|--------------------------|----------------------------------|------------------------------------|--------------------------------|-----------------------------------| |
|
| Korean | KMMLU | **53.26** (1st) | 45.30 | 45.28 | 42.17 | <u>45.66</u> | 41.66 | 31.45 | 38.49 | 41.50 | |
|
| | KMMLU-hard | **29.46** (1st) | 23.17 | 20.78 | 19.25 | <u>24.78</u> | 20.49 | 17.86 | 19.83 | 20.61 | |
|
| | KoBEST | **83.40** (1st) | 79.05 | 80.13 | <u>81.67</u> | 78.51 | 67.56 | 63.77 | 72.99 | 73.26 | |
|
| | Belebele | **57.99** (1st) | 40.97 | 45.11 | 49.40 | <u>54.85</u> | 54.70 | 40.31 | 53.17 | 48.68 | |
|
| | CSATQA | <u>43.32</u> (2nd) | 40.11 | 34.76 | 39.57 | **45.45** | 36.90 | 27.27 | 32.62 | 34.22 | |
|
| English | MMLU | 66.64 (3rd) | 65.27 | 64.32 | 63.63 | **74.26** | <u>68.26</u> | 62.04 | 63.25 | 65.30 | |
|
| | MMLU-Pro | **43.05** (1st) | 40.73 | 38.90 | 32.79 | <u>42.5</u> | 40.92 | 33.49 | 37.11 | 30.25 | |
|
| | GSM8K | **80.52** (1st) | 65.96 | <u>80.06</u> | 56.18 | 75.74 | 75.82 | 49.66 | 64.14 | 69.22 | |
|
- The *highest* *scores* are in **bold** form, and the *second*\-*highest* *scores* are <u>underlined</u>. |
|
|
|
**Evaluation Protocol** |
|
For easy reproduction of our evaluation results, we list the evaluation tools and settings used below: |
|
|
|
| | Evaluation setting | Metric | Evaluation tool | |
|
|------------|--------------------|-------------------------------------|-----------------| |
|
| KMMLU | 5-shot | macro\_avg / exact\_match | lm-eval-harness | |
|
| KMMLU Hard | 5-shot | macro\_avg / exact\_match | lm-eval-harness | |
|
| KoBEST | 5-shot | macro\_avg / f1 | lm-eval-harness | |
|
| Belebele | 0-shot | acc | lm-eval-harness | |
|
| CSATQA | 0-shot | acc\_norm | lm-eval-harness | |
|
| MMLU | 5-shot | macro\_avg / acc | lm-eval-harness | |
|
| MMLU Pro | 5-shot | macro\_avg / exact\_match | lm-eval-harness | |
|
| GSM8K | 5-shot | acc, exact\_match & strict\_extract | lm-eval-harness | |
|
|
|
## Quickstart |
|
|
|
This model requires `transformers >= 4.43.0`. |
|
|
|
```python |
|
from transformers import AutoModelForCausalLM, AutoTokenizer, TextStreamer |
|
|
|
tokenizer = AutoTokenizer.from_pretrained('dnotitia/Llama-DNA-1.0-8B-Instruct') |
|
model = AutoModelForCausalLM.from_pretrained('dnotitia/Llama-DNA-1.0-8B-Instruct', device_map='auto') |
|
streamer = TextStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True) |
|
|
|
conversation = [ |
|
{"role": "system", "content": "You are a helpful assistant, Dnotitia DNA."}, |
|
{"role": "user", "content": "λμ μ΄λ¦μ?"}, |
|
] |
|
inputs = tokenizer.apply_chat_template(conversation, |
|
add_generation_prompt=True, |
|
return_dict=True, |
|
return_tensors="pt").to(model.device) |
|
_ = model.generate(**inputs, streamer=streamer) |
|
``` |
|
|
|
## Limitations |
|
|
|
While DNA 1.0 8B Instruct demonstrates strong performance, users should be aware of the following limitations: |
|
|
|
- The model may occasionally generate biased or inappropriate content |
|
- Responses are based on training data and may not reflect current information |
|
- The model may sometimes produce factually incorrect or inconsistent answers |
|
- Performance may vary depending on the complexity and domain of the task |
|
- Generated content should be reviewed for accuracy and appropriateness |
|
|
|
## License |
|
|
|
This model is released under CC BY-NC 4.0 license. For commercial usage inquiries, please [Contact us](https://www.dnotitia.com/contact/post-form). |
|
|
|
## Appendix |
|
|
|
- KMMLU scores comparison chart: |
|
<img src="assets/comparison-chart.png" width="100%" style="margin: 40px auto;"> |
|
|
|
- DNA 1.0 8B Instruct model architecture <sup>[1]</sup>: |
|
<img src="assets/model-architecture.png" width="500" style="margin: 40px auto;"> |
|
|
|
[1]: <https://www.linkedin.com/posts/sebastianraschka_the-llama-32-1b-and-3b-models-are-my-favorite-activity-7248317830943686656-yyYD/> |
|
|
|
- The median percentage of modelβs weight difference between before and after the merge (our SFT model + Llama 3.1 8B Instruct): |
|
<img src="assets/ours-vs-merged.png" width="100%" style="margin: 40px auto;"> |
|
|
|
## Citation |
|
|
|
If you use or discuss this model in your academic research, please cite the project to help spread awareness: |
|
|
|
``` |
|
@misc{lee2025dna10technicalreport, |
|
title={DNA 1.0 Technical Report}, |
|
author={Jungyup Lee and Jemin Kim and Sang Park and SeungJae Lee}, |
|
year={2025}, |
|
eprint={2501.10648}, |
|
archivePrefix={arXiv}, |
|
primaryClass={cs.CL}, |
|
url={https://arxiv.org/abs/2501.10648}, |
|
} |
|
``` |
|
|
|
|
