Instructions to use nowdoor/Qwen3.8-27B-Inspect-v01 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use nowdoor/Qwen3.8-27B-Inspect-v01 with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("image-text-to-text", model="nowdoor/Qwen3.8-27B-Inspect-v01") 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("nowdoor/Qwen3.8-27B-Inspect-v01") model = AutoModelForMultimodalLM.from_pretrained("nowdoor/Qwen3.8-27B-Inspect-v01", 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 nowdoor/Qwen3.8-27B-Inspect-v01 with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "nowdoor/Qwen3.8-27B-Inspect-v01" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "nowdoor/Qwen3.8-27B-Inspect-v01", "messages": [ { "role": "user", "content": [ { "type": "text", "text": "Describe this image in one sentence." }, { "type": "image_url", "image_url": { "url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg" } } ] } ] }'Use Docker
docker model run hf.co/nowdoor/Qwen3.8-27B-Inspect-v01
- SGLang
How to use nowdoor/Qwen3.8-27B-Inspect-v01 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 "nowdoor/Qwen3.8-27B-Inspect-v01" \ --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": "nowdoor/Qwen3.8-27B-Inspect-v01", "messages": [ { "role": "user", "content": [ { "type": "text", "text": "Describe this image in one sentence." }, { "type": "image_url", "image_url": { "url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg" } } ] } ] }'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 "nowdoor/Qwen3.8-27B-Inspect-v01" \ --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": "nowdoor/Qwen3.8-27B-Inspect-v01", "messages": [ { "role": "user", "content": [ { "type": "text", "text": "Describe this image in one sentence." }, { "type": "image_url", "image_url": { "url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg" } } ] } ] }' - Docker Model Runner
How to use nowdoor/Qwen3.8-27B-Inspect-v01 with Docker Model Runner:
docker model run hf.co/nowdoor/Qwen3.8-27B-Inspect-v01
Qwen3.8 27b inspect v01
Qwen/Qwen3.8-27B์ ์ธ์ด ๋ชจ๋ธ ์ ์ฒด ํ๋ผ๋ฏธํฐ๋ฅผ ํ๊ตญ์ด ์์ค๋ฌผ ์์ ์ ๊ฒยท์ ๋ฐ์์ ์ง๋จ ๋ณด๊ณ ์ ์๋ฏธ ๊ฒํ ์์
์ ์ง๋ ๋ฏธ์ธ์กฐ์ ํ ๋ชจ๋ธ์
๋๋ค. ๋น์ ์ธ์ฝ๋์ ๋ณด์กฐ MTP ํ
์๋ ์๋ณธ ๋ชจ๋ธ์์ ๋ณด์กดํ ๋ค ์ต์ข
VLM ์ฒดํฌํฌ์ธํธ์ ๋ณํฉํ์ต๋๋ค.
์ด ์ ์ฅ์์๋ ๋ณํฉ๋ ์ ์ฒด ๋ชจ๋ธ ๊ฐ์ค์น, ํ ํฌ๋์ด์ ยทํ๋ก์ธ์ ์ค์ , ๋ผ์ด์ ์ค์ ๋ชจ๋ธ ์นด๋๋ง ํฌํจ๋ฉ๋๋ค. ํ์ตยทํ๊ฐ ๋ฐ์ดํฐ์ ์คํ ๋ก๊ทธ๋ ํฌํจํ์ง ์์ผ๋ฉฐ, ๋ชจ๋ธ ์นด๋์ ํ์ต ์ค์ ๊ณผ ๋ ๋ฆฝ ํ๊ฐ ๊ฒฐ๊ณผ์ ์ง๊ณ๋ง ์์ฝํ์ต๋๋ค.
Intended Use
- ์์ค๋ฌผ ๋ณด๊ณ ์์ ์ง์นจ ์ค์ ์ฌ๋ถ์ ๋ํ ๋ณด์กฐ ๊ฒํ
- ๊ทผ๊ฑฐ ์กฐํญ, ์ ์ฉ์กฐ๊ฑด, ๋ณด๊ณ ์ ์ธ์ฉ, ๋ถ์กฑํ ์ ๊ณผ ๋ณด์๋ฐฉ๋ฒ์ ํฌํจํ JSON ์์ฑ
- ๊ฒํ ์ด์ ์์ฑ๊ณผ ํ์ง๊ด๋ฆฌ ์ง์
์์ค๋ฌผ ์์ ๋ฑ๊ธ ๊ฒฐ์ , ๋ฒ์ ์ ํฉ์ฑ ํ๋จ ๋๋ ํ์ฅ ๊ธฐ์ ์์ ์ต์ข ๊ฒฐ์ ์ ๋์ฒดํ๋๋ก ๋ง๋ ๋ชจ๋ธ์ด ์๋๋๋ค. ์ค์ ์ฌ์ฉ ์ ์๊ฒฉ์ ๊ฐ์ถ ๊ธฐ์ ์๊ฐ ์ ๋ ฅ ์๋ฌธ, ์ ์ฉ ์ง์นจ๊ณผ ๋ชจ๋ธ ๊ฒฐ๋ก ์ ํ์ธํด์ผ ํฉ๋๋ค.
Training
| ํญ๋ชฉ | ๊ฐ |
|---|---|
| ์๋ณธ ๋ชจ๋ธ | Qwen/Qwen3.8-27B |
| ๋ฐฉ๋ฒ | ์ธ์ด ๋ชจ๋ธ ์ ์ฒด ํ๋ผ๋ฏธํฐ SFT, assistant-only loss |
| ํ๋ผ๋ฏธํฐ ์ | 27B |
| ํ์ต/๋ด๋ถ ํ๊ฐ ์ฌ๋ก | 1,960 / 40 |
| epoch | 1 |
| ์ต๋ ๊ธธ์ด | 18,496 tokens |
| precision | BF16, TF32 |
ํ์ต ์๋ฃ๋ 2,000๊ฐ์ ๋ก์ปฌ ์์ค๋ฌผ ๊ฒํ ๋ ์ฝ๋๋ก ๊ตฌ์ฑ๋์ผ๋ฉฐ ๊ฐ์ธ์ ๋ณด๊ฐ ๋น์๋ณํ๋ ๋ณด๊ณ ์ ๋ฐ์ท์ ํฉ์ฑ ์ฌ๋ก๋ฅผ ํฌํจํฉ๋๋ค. ๋ฐ์ดํฐ ์๋ฌธ์ ์ด ์ ์ฅ์์์ ๋ฐฐํฌํ์ง ์์ต๋๋ค. ๊ณต๊ฐ ๋ฐฐํฌ ์ ์ ๋ฐ์ดํฐ์ ํ์ ๋ชจ๋ธ์ ๋ํ ๊ถ๋ฆฌ๋ฅผ ๋ณ๋๋ก ํ์ธํด์ผ ํฉ๋๋ค.
ํ์ต ๋ฐ ๋ณํฉ ๊ฒ์ฆ ๋ก๊ทธ๋ ๋ก์ปฌ ๊ฒ์ฆ์ฉ์ผ๋ก๋ง ๋ณด๊ดํ๋ฉฐ ์ด ์ ์ฅ์์๋ ๋ฐฐํฌํ์ง ์์ต๋๋ค.
Evaluation
Facility Review
ํ์ต 2,000๊ฑด๊ณผ ์
๋ ฅยท์ถ๋ ฅยท์ฌ๋ก ID์ ์ ํ ์ค๋ณต์ด ์๋ ๋
๋ฆฝ ํฉ์ฑ ํ
์คํธ 50๊ฑด์ ์ฌ์ฉํ์ต๋๋ค. ๋ ๋ชจ๋ธ ๋ชจ๋ ๋์ผ ํ๋กฌํํธ, greedy decoding, temperature=0, enable_thinking=False๋ก ํ๊ฐํ์ต๋๋ค.
| ์งํ | Base | Fine-tuned |
|---|---|---|
| 6๋จ๊ณ ์ข ํฉํ์ ์ ํ๋ | 32/50 (64.0%) | 36/50 (72.0%) |
| ์ผ๋ฐ์ง์นจ 25๊ฑด | 60.0% | 64.0% |
| ๊ต๋ ์ธ๋ถ์ง์นจ 25๊ฑด | 68.0% | 80.0% |
| JSON ํ์ฑ ์ฑ๊ณต | 0/50 | 50/50 |
| ํ์ ํ๋ ์คํค๋ง ์ถฉ์กฑ | 0/50 | 49/50 |
| ์กฐํญยทํ์ด์งยท์ฆ๊ฑฐ ์ถ์ ์ฑ | 0/50 | 49/50 |
ํ์ ์ ํ๋ ์ฐจ์ด๋ +8.0%p์ง๋ง paired bootstrap 95% CI๊ฐ -8.0~+24.0%p์ด๊ณ McNemar exact p=0.4545์ด๋ฏ๋ก, ์ด ํ๋ณธ๋ง์ผ๋ก ์๋ฏธ ํ์ ์ฑ๋ฅ ํฅ์์ ํ์ ํ ์ ์์ต๋๋ค. ๊ตฌ์กฐํ ์ถ๋ ฅ์ ๋ช
ํํ ๊ฐ์ ๋์ต๋๋ค.
Fine-tuned ๋ชจ๋ธ์ ์ ๋ต์ ๋จ์ผ ๊ธฐ์ค์ ๋๋ถ๋ถ 2~3๊ฐ ํ์ ๊ธฐ์ค์ผ๋ก ๋ถํดํ๊ณ ์์ฒด ๊ธฐ์ค ID๋ฅผ ์์ฑํ์ต๋๋ค. ๋ฐ๋ผ์ ์ ํํ ํ ๊ฐ์ ๊ธฐ์ค๊ณผ ๋ฏธ๋ฆฌ ์ง์ ๋ ID๊ฐ ํ์ํ ์์คํ
์ JSON Schema ์ ์ฝ ๋๋ ์ถ๊ฐ ํ์ต์ด ํ์ํฉ๋๋ค. ๋ํ ๋ฏธํก ์ฌํ์จ์ 100.0%์์ 71.4%, ๋น์ ์ฉ์ 100.0%์์ 50.0%๋ก ๋ฎ์์ ธ ํ์ ์ ์ํํ๋ ๊ฒฝํฅ์ด ๊ด์ฐฐ๋์ต๋๋ค.
AIME 2024+2025
์ด 60๋ฌธํญ์ thinking mode, medium reasoning effort, greedy decoding, ์ต๋ 8,192 ์์ฑ ํ ํฐ์ผ๋ก ํ๊ฐํ์ต๋๋ค.
| ์งํ | Base | Fine-tuned |
|---|---|---|
| ์ถ์ถ ๋ต์ ์ ํ๋ | 47/60 (78.33%) | 46/60 (76.67%) |
| ๋ณํ | -1.67%p |
McNemar exact p=1.0์ด๋ฉฐ paired bootstrap ์ ๋ขฐ๊ตฌ๊ฐ์ 0์ ํฌํจํฉ๋๋ค. ์์ค๋ฌผ ๋๋ฉ์ธ ๋ฏธ์ธ์กฐ์ ํ AIME ์ฑ๋ฅ ํฅ์์ ํ์ธ๋์ง ์์์ต๋๋ค. ์์ฑ ๋ฏธ์๋ฃ๋ฅผ ์ค๋ต ์ฒ๋ฆฌํ๋ ์๊ฒฉํ ๋ฏผ๊ฐ๋ ๋ถ์์์๋ Base 30/60, Fine-tuned 35/60์ด์์ต๋๋ค.
Usage
ํ๊ฐ์ ๊ฐ์ ์กฐ๊ฑด์์๋ thinking์ ๋นํ์ฑํํ๊ณ greedy decoding์ ์ฌ์ฉํฉ๋๋ค.
import torch
from transformers import AutoModelForImageTextToText, AutoProcessor
model_id = "nowdoor/Qwen3.8-27B-Inspect-v01"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForImageTextToText.from_pretrained(
model_id,
torch_dtype=torch.bfloat16,
device_map="auto",
)
messages = [
{
"role": "system",
"content": (
"๋๋ ์์ค๋ฌผ ์์ ์ ๊ฒยท์ ๋ฐ์์ ์ง๋จ ๋ณด๊ณ ์๋ฅผ ์ผ๋ฐ์ง์นจ์ ๋ฐ๋ผ "
"์๋ฏธ ์ค์ฌ์ผ๋ก ๊ฒํ ํ๋ ์ ๋ฌธ๊ฐ๋ค. ์
๋ ฅ์ ์๋ ์ฌ์ค์ ์ถ์ ํ์ง ์๋๋ค."
),
},
{
"role": "user",
"content": "๋ณด๊ณ ์ ๋ฐ์ท์ ๊ฒํ ๊ธฐ์ค์ ์ฌ๊ธฐ์ ์
๋ ฅํฉ๋๋ค.",
},
]
prompt = processor.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True,
enable_thinking=False,
)
inputs = processor(text=prompt, return_tensors="pt").to(model.device)
with torch.inference_mode():
output = model.generate(**inputs, max_new_tokens=2048, do_sample=False)
new_tokens = output[:, inputs["input_ids"].shape[1]:]
print(processor.batch_decode(new_tokens, skip_special_tokens=True)[0])
๊ฒ์ฆ ํ๊ฒฝ์ Transformers 5.15.0 ์ง์ ์ถ๋ก ๊ณผ vLLM 0.17.1์ ๋๋ค. ๋ชจ๋ธ์ ๋ค์ดํฐ๋ธ ์ต๋ ๋ฌธ๋งฅ์ ์๋ณธ ์ค์ ์ ์ ์งํ์ง๋ง, ๋ณธ ๋ฏธ์ธ์กฐ์ ๋ฐ์ดํฐ๋ ์ต๋ 18,496ํ ํฐ๊น์ง๋ง ํฌํจํ์ต๋๋ค.
Limitations
- 50๊ฑด ์์ค๋ฌผ ํ๊ฐ๋ ํฉ์ฑ ๋ ๋ฆฝ ์ฌ๋ก์ด๋ฉฐ ์ค์ ์ ์ฒด ๋ณด๊ณ ์์ ๋ํ ์ธ๋ถ ํ๋น์ฑ์ ๋ณด์ฅํ์ง ์์ต๋๋ค.
ํ๋จ๋ณด๋ฅ์ฌ๋ก๋ ๋ ๋ชจ๋ธ ๋ชจ๋ ๋งํ์ง ๋ชปํ์ต๋๋ค.- ์๊ฒฉํ
๋ฏธํก๋๋๋น์ ์ฉํ์ ์ ์ํํ ์ ์์ต๋๋ค. - ๊ธฐ์ค ID์ ๊ฒฐ๊ณผ ํญ๋ชฉ ์๊ฐ ํธ์ถ์๊ฐ ๊ธฐ๋ํ๋ ๊ณ์ฝ๊ณผ ๋ค๋ฅผ ์ ์์ต๋๋ค.
- ๋ณด๊ณ ์์ ์๋ ์ฌ์ค์ด๋ ๊ณผ๋ํ ๋ณด์ ๊ทผ๊ฑฐ๋ฅผ ์์ฑํ ๊ฐ๋ฅ์ฑ์ด ์์ผ๋ฏ๋ก ์๋ฌธ ๋์กฐ๊ฐ ํ์ํฉ๋๋ค.
- ๋น์ ์ธ์ฝ๋๋ ๋ณด์กด๋์ง๋ง ๋ณํฉ ํ ์ด๋ฏธ์งยท์์ ์ ๋ ฅ์ ๋ํ ๋ณ๋ ๋๋ฉ์ธ ํ๊ฐ๋ ์ํํ์ง ์์์ต๋๋ค.
- AIME์์๋ Base๋ณด๋ค 1๋ฌธํญ ๋ฎ์์ผ๋ฏ๋ก ์ผ๋ฐ ์ํ ์ฑ๋ฅ ํฅ์์ ๊ธฐ๋ํด์๋ ์ ๋ฉ๋๋ค.
License
Apache License 2.0. ์๋ณธ ๋ชจ๋ธ์ ๋ผ์ด์ ์ค์ ์ฌ์ฉ ์กฐ๊ฑด๋ ํจ๊ป ํ์ธํ์ญ์์ค.
- Downloads last month
- 9
Model tree for nowdoor/Qwen3.8-27B-Inspect-v01
Base model
Qwen/Qwen3.8-27BEvaluation results
- Six-level judgment accuracy on AutoCheck independent synthetic testtest set self-reported0.720
- Extracted-answer accuracy on AIME 2024 and 2025test set self-reported0.767