Instructions to use Anserwise/AWAXIS-Hybrid-28B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use Anserwise/AWAXIS-Hybrid-28B with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-generation", model="Anserwise/AWAXIS-Hybrid-28B") messages = [ { "role": "user", "content": [ {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"}, {"type": "text", "text": "What animal is on the candy?"} ] }, ] pipe(text=messages)# pip install -U transformers accelerate # Load model directly from transformers import AutoProcessor, AutoModelForMultimodalLM processor = AutoProcessor.from_pretrained("Anserwise/AWAXIS-Hybrid-28B") model = AutoModelForMultimodalLM.from_pretrained("Anserwise/AWAXIS-Hybrid-28B", device_map="auto") messages = [ { "role": "user", "content": [ {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"}, {"type": "text", "text": "What animal is on the candy?"} ] }, ] inputs = processor.apply_chat_template( messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt", ).to(model.device) outputs = model.generate(**inputs, max_new_tokens=256) print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:])) - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- vLLM
How to use Anserwise/AWAXIS-Hybrid-28B with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "Anserwise/AWAXIS-Hybrid-28B" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "Anserwise/AWAXIS-Hybrid-28B", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }'Use Docker
docker model run hf.co/Anserwise/AWAXIS-Hybrid-28B
- SGLang
How to use Anserwise/AWAXIS-Hybrid-28B with SGLang:
Install from pip and serve model
# Install SGLang from pip: pip install sglang # Start the SGLang server: python3 -m sglang.launch_server \ --model-path "Anserwise/AWAXIS-Hybrid-28B" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "Anserwise/AWAXIS-Hybrid-28B", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }'Use Docker images
docker run --gpus all \ --shm-size 32g \ -p 30000:30000 \ -v ~/.cache/huggingface:/root/.cache/huggingface \ --env "HF_TOKEN=<secret>" \ --ipc=host \ lmsysorg/sglang:latest \ python3 -m sglang.launch_server \ --model-path "Anserwise/AWAXIS-Hybrid-28B" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "Anserwise/AWAXIS-Hybrid-28B", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }' - Docker Model Runner
How to use Anserwise/AWAXIS-Hybrid-28B with Docker Model Runner:
docker model run hf.co/Anserwise/AWAXIS-Hybrid-28B
AWAXIS-Hybrid-28B
AWAXIS-Think ร Darwin Platform Hybrid ๋ชจ๋ธ
AWAXIS-Think-28B๋ฅผ ๊ธฐ๋ฐ์ผ๋ก Darwin Platform ํ๊ตญ์ด ๊ฐ์ค์น๋ฅผ Smart MRI Layer-wise ๋จธ์ง๋ก ๊ฒฐํฉํ ํ๊ตญ์ด LLM์ ๋๋ค.
๋ณธ ๋ชจ๋ธ์ Anserwise์์ ์ ์ยท๊ณต๊ฐํ ํ๊ตญ์ด LLM์ ๋๋ค.
- ๐งฌ ์๋ฒ์ง:
Anserwise/AWAXIS-Think-28B - ๐งฌ ์ด๋จธ๋:
FINAL-Bench/Darwin-28B-KR
๐งฌ ๋ชจ๋ธ ์ค๋ช โ ๋ถ๋ชจ ๋ชจ๋ธ ๊ต๋ฐฐยท์งํ (Darwin ๋ฐฉ์)
๋ณธ ๋ชจ๋ธ์ ๋ ๋ถ๋ชจ LLM์ ๊ฐ์ค์น๋ฅผ layer ๋จ์๋ก ๊ต๋ฐฐยท๊ฒฐํฉํ๋ Darwin Platform ์งํ์ ๋จธ์ง ๊ธฐ๋ฒ (Smart MRI Layer-wise)์ผ๋ก ์ ์๋์ต๋๋ค. ์ผ๋ฐ SFT/CPT ํ์ต์ด ์๋ ๊ฐ์ค์น ํฉ์ฑ ์งํ ๋ฐฉ์์ด๋ผ๋ ์ ์ด ํต์ฌ์ ๋๋ค.
| ๊ตฌ๊ฐ | ์ด๋จธ๋ ์ฑํ๋ฅ | ์๋ |
|---|---|---|
| Embed / LM-head | 50% | ์ถ๋ ฅ ํต๋ก ๊ท ํ |
| Norm | 30% | ์์ ์ฑ |
| Visual encoder | 0% | ์๋ฒ์ง 100% ๋ณด์กด |
| Layers 0~15 (์ด๊ธฐ) | 40% | ํ๊ตญ์ด ํ๋ฉด ํจํด ํก์ |
| Layers 16~50 (์ค๊ธฐ) | 0% | ์ถ๋ก ๋ฅ๋ ฅ ๋ณด์กด |
| Layers 51~63 (ํ๊ธฐ) | 70% | ๋๋ฉ์ธ ์ง์ ์ฑํ |
๊ฐ ๋ถ๋ชจ์ ๊ฐ์ ์ด ์ด๋ layer์ ์ ์ฅ๋๋์ง ๋ถ์ํ ํ layer-wise๋ก ๋ถ๋ชจ ๊ฐ์ค์น ๋น์จ์ ๋ค๋ฅด๊ฒ ์ ์ฉํ์ฌ ๋ ๋ชจ๋ธ์ ๊ฐ์ ๋ง ์ ํ์ ์ผ๋ก ๊ฒฐํฉํ๋ ๋ฐฉ์์ ๋๋ค.
๐ ๋ฐ์ดํฐ์ ํ์ฉ
๋ณธ ๋ชจ๋ธ์ ๊ฐ์ค์น ๋จธ์ง(Weight Merge) ์ฐ๋ฌผ์ด๋ฏ๋ก ์ถ๊ฐ SFT/CPT ํ์ต์ ์ํํ์ง ์์์ต๋๋ค. ๋์ ๋ ๋ถ๋ชจ ๋ชจ๋ธ์ด ์ฌ์ ์ ํ์ตํ ๊ด๋ฒ์ํ ํ๊ตญ์ด ๋ฐ์ดํฐ์ ์ ๊ฐ์ ์ layer ๋จ์๋ก ํก์ํฉ๋๋ค. ๋ถ๋ชจ ๋ชจ๋ธ๋ค์ด ํ์ตํ ์ฃผ์ ํ๊ตญ์ด ๋ฐ์ดํฐ์ ์ ๋ค์๊ณผ ๊ฐ์ต๋๋ค:
- ์ผ๋ฐ Instruction: kai-sft / kai-combined ์๋ฆฌ์ฆ (ํ๊ตญ์ด ๋ค์ ๋๋ฉ์ธ instruction-following)
- KMMLU-Pro: ํ๊ตญ ๋๋ฉ์ธ ์ง์ (์ญ์ฌยท๋ฒ๋ฅ ยท์๋ฃยท๊ณผํยท๊ณตํ ๋ฑ)
- CLIcK: ํ๊ตญ ๋ฌธํยท์์ (์ญ์ฌยท์ ํตยท์ฌํ ๊ท๋ฒ)
- HAERAE: ํ๊ตญ์ด ํ๋ฉด ํจํด (์ธ์ดํยท์ผ๋ฐ์์ยท์ญ์ฌ)
- KOBEST: ํ๊ตญ์ด reasoning (HellaSwagยทCOPAยทBoolQ)
- Com2-main(ko): ํ๊ตญ์ด commonsense (์ฌํ์ ์ถ๋ก ยท์๋ ํ์ )
- MuSR(Ko): ํ๊ตญ์ด ๋ค๋จ๊ณ ์ถ๋ก (Murder MysteryยทObject Placements ๋ฑ)
๊ฐ์ค์น ๋จธ์ง ๋จ๊ณ์์ ๊ฐ ๋ถ๋ชจ์ ํ์ต ๊ฒฐ๊ณผ๋ฅผ layer ๋น์จ๋ก ๋ณด์กดยท๊ฒฐํฉํ์ฌ, ๋จ์ผ SFT๋ณด๋ค catastrophic forgetting ์ํ์ด ๋ฎ๊ณ ๋ถ๋ชจ ๊ฐ์ ์์ค์ด ์ ์ ๊ฒ์ด ํน์ง์ ๋๋ค.
๐ ํ๊ฐ ๊ฒฐ๊ณผ
1) K-AI ๋ฆฌ๋๋ณด๋ ๊ธฐ์ค (5๊ณผ๋ชฉ)
KMMLU-Pro / CLIcK / HLE(Ko) / MuSR(Ko) / Com2-main(ko)
- ์ธ๋ถ ๋ชจ๋ธ: K-AI ๋ฆฌ๋๋ณด๋(leaderboard.aihub.or.kr) ์ค์ธก๊ฐ
- ๋ณธ ์๋ฆฌ์ฆ: ์์ฒด mirror eval(100๋ฌธํญ) ร Rogue-28B-MIX ๊ธฐ์ค ratio ํ์ฐ ์ถ์ ๊ฐ
| Model | KMMLU-Pro | CLIcK | HLE(Ko) | MuSR(Ko) | Com2-main(ko) | Sum | Macro |
|---|---|---|---|---|---|---|---|
| Hybrid (์์) | 0.674 | 0.787 | 0.07 | 0.611 | 0.657 | 2.799 | 0.560 |
| AWAXIS-Hybrid-28B โญ (์ด ๋ชจ๋ธ, ์์) | 0.674 | 0.787 | 0.07 | 0.611 | 0.657 | 2.799 | 0.560 |
| Rogue-28B-MIX (์ค์ธก) | 0.666 | 0.797 | 0.07 | 0.611 | 0.650 | 2.794 | 0.559 |
| Warecube-KO-27B-v3 (์ค์ธก) | 0.668 | 0.799 | 0.07 | 0.584 | 0.638 | 2.756 | 0.551 |
| AWAXIS-Think-28B (์ค์ธก) | 0.603 | 0.770 | 0.06 | 0.591 | 0.632 | 2.651 | 0.530 |
| KR-Pro (์์) | 0.643 | 0.661 | 0.07 | 0.585 | 0.650 | 2.609 | 0.522 |
| KR-Plus (์์) | 0.643 | 0.703 | 0.07 | 0.532 | 0.657 | 2.605 | 0.521 |
HLE(Ko)๋ 28B๊ธ ๊ณตํต ์ฝ์ (๋์ด๋ ๋งค์ฐ ๋์).
2) ์ข ํฉ ํ๊ตญ์ด ๋ฅ๋ ฅ (10๊ณผ๋ชฉ mirror eval)
CLIcK + KMMLU(history/law/health) + HAERAE(gk/hist/ling) + KOBEST(hella/copa/boolq)
| Model | CLIcK | KMMLU ํ๊ท | HAERAE ํ๊ท | KOBEST ํ๊ท | Sum (10๊ณผ๋ชฉ) | Macro |
|---|---|---|---|---|---|---|
| AWAXIS-Hybrid-28B โญ (์ด ๋ชจ๋ธ) | 0.83 | 0.530 | 0.813 | 0.967 | 7.760 | 0.7760 |
| Rogue-28B-MIX | 0.83 | 0.513 | 0.807 | 0.967 | 7.690 | 0.7690 |
์ฌ์ฉ ๋ฐฉ๋ฒ
from transformers import AutoTokenizer, AutoModelForCausalLM
tokenizer = AutoTokenizer.from_pretrained("Anserwise/AWAXIS-Hybrid-28B", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
"Anserwise/AWAXIS-Hybrid-28B",
torch_dtype="bfloat16",
device_map="auto",
trust_remote_code=True,
)
messages = [{"role": "user", "content": "ํ๊ตญ ์ญ์ฌ์์ ์์ง์๋์ด ๋ฐ์ํ ์๊ธฐ๋?"}]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True).to(model.device)
outputs = model.generate(inputs, max_new_tokens=256, temperature=0.0)
print(tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True))
๋ผ์ด์ ์ค
Apache 2.0 (๋ถ๋ชจ ๋ชจ๋ธ ๋ผ์ด์ ์ค ๊ณ์น)
2026-04-30
ํ์ต ๋ฐ์ดํฐ์
AI Hub(๊ณผํ๊ธฐ์ ์ ๋ณดํต์ ๋ถ ์ฐํ NIA ์ด์) ๊ณต๊ฐ ํ๊ตญ์ด ๋ง๋ญ์น๋ฅผ ํ์ฉํ์ฌ ํ์ตํ์์ต๋๋ค.
ํ๊ตญ์ด ๋ํ ๋ฐ์ดํฐ์ https://aihub.or.kr/aihubdata/data/view.do?dataSetSn=272
๋๊ท๋ชจ ์น๋ฐ์ดํฐ ๊ธฐ๋ฐ ํ๊ตญ์ด ๋ง๋ญ์น https://aihub.or.kr/aihubdata/data/view.do?dataSetSn=624
๋ฉํฐ์ธ์ ยท๋ํ ๊ด๋ จ ๋ง๋ญ์น ๋ฐ์ดํฐ https://aihub.or.kr/aihubdata/data/view.do?currMenu=511&topMenu=100&aihubDataSe=dataPckage&dataPckageSn=1
ํ๊ตญ์ด ์์ฑ ๊ธฐ๋ฐ ์์์ถ๋ก ๋ฐ์ดํฐ์ https://aihub.or.kr/aihubdata/data/view.do?dataSetSn=459
์ ๋ฌธ ์ํ์ง์ ๋ฐ์ดํฐ https://aihub.or.kr/aihubdata/data/view.do?dataSetSn=71874
์๋ฃยท๋ฒ๋ฅ ์ ๋ฌธ ์์ ๋ง๋ญ์น https://aihub.or.kr/aihubdata/data/view.do?dataSetSn=71487
- Downloads last month
- 9