1. Model Description
- KONI (KISTI Open Natural Intelligence) is a specialized large language model (LLM) developed by the Korea Institute of Science and Technology Information (KISTI). This model is specifically designed for science and technology, making it highly effective for tasks in these fields.
2. Key Features
- Specialized in Science and Technology: The model is explicitly trained on a vast and specialized corpus of scientific and technological data.
- Enhanced Performance: This version of KONI shows significantly improved performance compared to its initial release in December, 2023.
- Base Model: The base model for KONI-Llama3-8B-Instruct-20240729 is KONI-Llama3-8B-Merged-20240724, which is a merger of Meta-Llama-3-8B and KISTI-KONI/KONI-Llama3-8B-20240630
- Alignment: SFT (Supervised Fine-Tuning) and DPO (Direct Preference Optimization) are applied
3. Data
- Approximately 11k SFT data and 7k DPO data are used.
- SFT Data: The SFT data includes both internally generated data and publicly available data on Hugging Face, translated into Korean where necessary.
- DPO Data: The DPO data consists of translated and curated data from argilla/dpo-mix-7k.
4. Benchmark Results
Results in LogicKor* are as follows:
| Metric | Score |
|---|---|
| Reasoning | 6.57 |
| Math | 8.00 |
| Writing | 8.92 |
| Coding | 8.85 |
| Comprehension | 9.85 |
| Grammar | 7.07 |
| Single-turn | 8.42 |
| Multi-turn | 8.00 |
| Overall | 8.21 |
| Our model demonstrates the best performance among publicly available 8B models on the LogicKor leaderboard as of 2024.07.30. |
5. How to use the model
import transformers
import torch
model_id = "KISTI-KONI/KONI-Llama3-8B-Instruct-20240729"
pipeline = transformers.pipeline(
"text-generation",
model=model_id,
model_kwargs={"torch_dtype": torch.bfloat16},
device_map="auto",
)
pipeline.model.eval()
instruction = "안녕? 너는 누구야?"
messages = [
{"role": "user", "content": f"{instruction}"}
]
prompt = pipeline.tokenizer.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True
)
terminators = [
pipeline.tokenizer.eos_token_id,
pipeline.tokenizer.convert_tokens_to_ids("<|eot_id|>")
]
outputs = pipeline(
prompt,
max_new_tokens=2048,
eos_token_id=terminators,
do_sample=True,
temperature=0.7,
top_p=0.9
)
print(outputs[0]["generated_text"][len(prompt):])
안녕하세요! 저는 KONI, 여러분의 질문에 답하고 정보를 제공하는 인공지능입니다. 저는 어떤 정보를 제공해 드릴까요?
6. Citation
Language Model
@article{KISTI-KONI/KONI-Llama3-8B-Instruct-20240729,
title={KISTI-KONI/KONI-Llama3-8B-Instruct-20240729},
author={KISTI},
year={2024},
url={https://huggingface.co/KISTI-KONI/KONI-Llama3-8B-Instruct-20240729}
}
7. Contributors
- KISTI, Large-scale AI Research Group
8. Special Thanks
8. Acknowledgement
- This research was supported by Korea Institute of Science and Technology Information(KISTI).
- This work was supported by the National Supercomputing Center with supercomputing resources including technical support (KISTI).