DataVortexS-10.7B-dpo-v1.0

Our Team

Research & Engineering	Product Management
Kwangseok Yang	Seunghyun Choi
Jeongwon Choi	Hyoseok Choi

Model Details

Base Model

megastudy/M-SOLAR-10.7B-v1.3

Trained On

• OS: Ubuntu 22.04
• GPU: H100 80GB 4ea
• transformers: v4.36.2

Instruction format

It follows Alpaca (Chat) format.

E.g.

text = """\
### System:
당신은 사람들이 정보를 찾을 수 있도록 도와주는 인공지능 비서입니다.

### User:
대한민국의 수도는 어디야?

### Assistant:
대한민국의 수도는 서울입니다.

### User:
서울 인구는 총 몇 명이야?
"""

Model Benchmark

Ko LM Eval Harness

Task	0-shot	5-shot	10-shot	50-shot
kobest_boolq	0.867265	0.930834	0.938736	0.938023
kobest_copa	0.722438	0.792716	0.782842	0.805869
kobest_hellaswag	0.484781	0.480055	0.496734	0.501488
kobest_sentineg	0.759887	0.964735	0.964735	0.972291
Average	0.70859275	0.792085	0.79576175	0.80441775

Ko-LLM-Leaderboard

Average	Ko-ARC	Ko-HellaSwag	Ko-MMLU	Ko-TruthfulQA	Ko-CommonGen V2
57.92	56.91	65.81	53.81	58.77	54.31

Implementation Code

This model contains the chat_template instruction format.
You can use the code below.

from transformers import AutoModelForCausalLM, AutoTokenizer

device = "cuda" # the device to load the model onto

model = AutoModelForCausalLM.from_pretrained("Edentns/DataVortexS-10.7B-dpo-v1.0")
tokenizer = AutoTokenizer.from_pretrained("Edentns/DataVortexS-10.7B-dpo-v1.0")

messages = [
    {"role": "system", "content": "당신은 사람들이 정보를 찾을 수 있도록 도와주는 인공지능 비서입니다."},
    {"role": "user", "content": "대한민국의 수도는 어디야?"},
    {"role": "assistant", "content": "대한민국의 수도는 서울입니다."},
    {"role": "user", "content": "서울 인구는 총 몇 명이야?"}
]

encodeds = tokenizer.apply_chat_template(messages, return_tensors="pt")

model_inputs = encodeds.to(device)
model.to(device)

generated_ids = model.generate(model_inputs, max_new_tokens=1000, do_sample=True)
decoded = tokenizer.batch_decode(generated_ids)
print(decoded[0])

License

The model is licensed under the cc-by-nc-sa-4.0 license, which allows others to copy, modify, and share the work non-commercially, as long as they give appropriate credit and distribute any derivative works under the same license.