shokuba-probe-2b
่ทๅ ด็นๆคOSS shokuba-lens ใฎ่ฟฝๅ ็ขบ่ช๏ผprobe๏ผ็จใซtokimoaใใใกใคใณใใฅใผใใณใฐใใ2B VLMใงใใๅ็ใซๅฏพใใ็นๆค่ณชๅ๏ผไพ: ใใซใกใใๆช็็จ่ ใฎไบบๆฐ๏ผใซใ็ญใๆง้ ๅๅ็ญใง็ญใใพใใ4bit้ๅญๅ+ใขใใใฟๅๆขฑใง็ด1.7GBใApple Silicon๏ผMLX๏ผใงๅไฝใใพใใ
ๆฑ็จVLMใฏ็นๆค่ณชๅใซๅฏพใใฆใ่ฆ้ใใชใใๆฐใ้ใใใๅพๅใใใ๏ผๅฎๆธฌใงprecision 0.57๏ผใ็นๆคใใผใฟใงใฎๅญฆ็ฟใซใใใใใ่งฃๆถใใฆใใพใใไปๅพใๅญฆ็ฟใ้ใญใฆๅฏพๅฟใใกใคใณใจๆฑๅๆง่ฝใๅบใใฆใใใพใ๏ผใขใใซๅใฏๆฎใ็ฝฎใใงๆดๆฐใใพใ๏ผใ
ๅฏพๅฟใใฆใใ็นๆค่ณชๅ๏ผ12็จฎ๏ผ
ใใซใกใใ็็จ / ๅฎๅ จใในใ็็จ / ๅฎๅ จๅธฏ๏ผใใซใใผใใน๏ผ / ็ซใป็ ใฎๆ็ก / ๆถ็ซๅจใฎๆฌๆฐ / ่ปขๅ่ ใฎๆ็ก / ใในใฏ็็จ / ใใขใใใ็็จ๏ผๅจๆฟ๏ผ / ๆ่ข็็จ๏ผๅจๆฟ๏ผ / ใใฉใผใฏใชใใใฎๅฐๆฐ๏ผๅๅบซใป็ฉๆต๏ผ / ใใฌใใใใฉใใฏใฎๅฐๆฐ๏ผๅๅบซใป็ฉๆต๏ผ / ใซใดใใฌใใใฎๆ็ก๏ผๅๅบซใป็ฉๆต๏ผ
ๅฎๆธฌ๏ผๅญฆ็ฟใจๅ้ขใใใในใใใผใฟใป็ปๅๅไฝ๏ผ
| ็นๆค่ณชๅ | ใในใ | ใใผใน๏ผQwen3-VL-2B 4bit๏ผ | shokuba-probe-2b |
|---|---|---|---|
| ใใซใกใใๆช็็จ่ ใฎๆๆ | SHEL5K 400ๆ | F1 0.520 | F1 0.860 |
| ๅฎๅ จใในใๆช็็จ่ ใฎๆๆ | 55ๆ | F1 0.415 | F1 0.897 |
| ๅฎๅ จๅธฏใฎ็ขบ่ช | 150ๆ | F1 0.650 | F1 0.878 |
| ใในใฏ็็จ็ถๆณ | 236ๆ๏ผ่ก้ ญ+ๅจๆฟ๏ผ | ไธ่ด็ 0.216 | ไธ่ด็ 0.636 |
| ใใขใใใๆช็็จใฎๆๆ | 150ๆ | F1 0.100 | F1 0.978 |
| ็ด ๆ๏ผๆ่ขใชใ๏ผใฎๆๆ | 150ๆ | F1 0.538 | F1 0.909 |
| ็ซใป็ ใฎๆ็ก | 200ๆ | ไธ่ด็ 0.660 | ไธ่ด็ 0.855 |
| ๆถ็ซๅจใฎๆฌๆฐ | 300ๆ | ไธ่ด็ 0.000 | ไธ่ด็ 0.930 |
| ่ปขๅ่ ใฎๆๆ | 450ๆ | ไธ่ด็ 0.000 | ไธ่ด็ 0.980 |
| ใใฉใผใฏใชใใใฎๅฐๆฐ | LOCO 150ๆ | ไธ่ด็ 0.000 | ไธ่ด็ 0.767 |
| ใใฌใใใใฉใใฏใฎๅฐๆฐ | LOCO 150ๆ | ไธ่ด็ 0.000 | ไธ่ด็ 0.707 |
| ใซใดใใฌใใใฎๆ็ก | LOCO 150ๆ | ไธ่ด็ 0.273 | ไธ่ด็ 0.860 |
ๆถ็ซๅจใป่ปขๅใปๅๅบซ3็จฎใฎใในใใฏๅฏพ่ฑกใๅใฃใฆใใชใ็ปๅใๅซใฟใพใ๏ผ่ชคๆค็ฅๅดใๆธฌๅฎ๏ผใ่ปขๅใฏๆฎๅฝฑ็ฐๅขใฎ็ฐใชใ2ใใผใฟใปใใ๏ผRoboflowใปCAUCAFall๏ผใงใๅๅบซ3็จฎใฏๅญฆ็ฟใจๆฎๅฝฑ็ฐๅขใ็ฐใชใ็ฉๆตๆ ็น๏ผLOCOๅ ฌๅผvalๅๅฒ๏ผใง่ฉไพกใใฆใใพใใ่ปขๅใฎๅญฆ็ฟใซใฏ็ฃ่ฆใซใกใฉๅฎๆ ๅ็ฑๆฅใฎ็ดใใใใๅงฟๅข๏ผๅบงใใใใป็ๅบงใป็ซใกไธใใใใ๏ผใฎ้ฐๆงใๅซใใฆใใใๅญฆ็ฟใจๅ้ขใใ็ฃ่ฆใซใกใฉๆ ๅ8ๆฌใป480ใใฌใผใ ใงใฎ่ปขๅ่ชคๆค็ฅใฏ0ไปถใงใใ
ๅญฆ็ฟใซไธๅไฝฟใฃใฆใใชใใใผใฟใปใใ๏ผSHWD ๅฎๅ จใใซใกใใๆคๅบ 150ๆใปๅญฆ็ฟ็ปๅใจใฎ้่คใชใใใใใทใฅๆค่จผๆธใฟ๏ผใฎใใซใกใใๅคๅฎใงใ F1 0.960 ใงใใๅ็ใ ใใงใชใๆจกๅผๅณใปๅณ่งฃใฎ่ชญใฟๅใใซใๅฏพๅฟใใฆใใพใใ
ไฝฟใๆน
shokuba-lensใใ:
python -m shokuba_lens.analyze --images site1.jpg \
--rules rules/construction_ppe.yaml \
--probe-model tokimoa/shokuba-probe-2b
ๅไฝใง:
from mlx_vlm import load, generate
from mlx_vlm.prompt_utils import apply_chat_template
from mlx_vlm.utils import load_config
repo = "tokimoa/shokuba-probe-2b" # ใญใผใซใซใซDLๆธใฟใชใใใฎใใน
model, processor = load(repo, adapter_path=repo) # ใขใใใฟๅๆขฑ
config = load_config(repo)
q = "ใใฎ็ปๅใซๅใฃใฆใใไบบใๆฐใใฆใใ ใใใๅ
จไฝใฎไบบๆฐใใใฎใใกใใซใกใใ๏ผไฟ่ญทๅธฝ๏ผใ็็จใใฆใใไบบๆฐใ็็จใใฆใใชใไบบๆฐใใไบบๆฐ:Nไบบใ็็จ:Nไบบใๆช็็จ:Nไบบใใฎๅฝขๅผใง็ญใใฆใใ ใใใ"
prompt = apply_chat_template(processor, config, q, num_images=1)
print(generate(model, processor, prompt, image=["site.jpg"], max_tokens=60, temperature=0.0).text)
่ณชๅๆใฏไธ่จใฎๅฝขๅผ๏ผใไบบๆฐ:Nไบบใ็็จ:Nไบบใๆช็็จ:Nไบบใ็ญใฎๅ็ญๅฝขๅผๆๅฎใคใ๏ผใๆใๅฎๅฎใใพใใ
ๆดๆฐๅฑฅๆญด
- 2026-08-04 (2): ใใซใใใผใฟใปใใๅญฆ็ฟ ็ฌฌ4ๅใ็ฃ่ฆใซใกใฉๅฎๆ ๅใฎ็ๅบงๅงฟๅขใซๅฏพใใ่ปขๅ่ชคๆค็ฅใ่งฃๆถ๏ผๅญฆ็ฟๅ้ขใฎๅฎๆ ๅ480ใใฌใผใ ใง่ชคๆค็ฅ0ไปถ๏ผใปใใซใกใใF1 0.843โ0.860ใปใใฌใใใใฉใใฏๆคๅบF1 0.833โ0.877ใปSHWDๆฑๅ0.954โ0.960
- 2026-08-04: ๅฏพๅฟ็นๆคใ9็จฎโ12็จฎใซๆกๅผต๏ผๅๅบซใป็ฉๆต3็จฎใๆฐ่จญใปใใซใใใผใฟใปใใๅญฆ็ฟ ็ฌฌ3ๅ๏ผใ่ปขๅๅญฆ็ฟใซ็ฃ่ฆใซใกใฉๅฎๆ ๅใฎ็ดใใใใๅงฟๅข้ฐๆงใ่ฟฝๅ
- 2026-08-03: ๅฏพๅฟ็นๆคใ8็จฎโ9็จฎใซๆกๅผต๏ผๅฎๅ จๅธฏใๆฐ่จญใปใใซใใใผใฟใปใใๅญฆ็ฟ ็ฌฌ2ๅใป11ใใผใฟใปใใ๏ผใๆชๅญฆ็ฟใใผใฟใปใใใธใฎๆฑๅ F1 0.967๏ผSHWD๏ผใปๆจกๅผๅณ่ชญใฟๅใๅฏพๅฟใป็ซ/็ 0.810โ0.865ใปๆ่ขF1 0.875โ0.944
- 2026-08-02: ๅฏพๅฟ็นๆคใ1็จฎโ8็จฎใซๆกๅผต๏ผใใซใใใผใฟใปใใๅญฆ็ฟ ็ฌฌ1ๅ๏ผใSHEL5K F1 0.837โ0.858
- 2026-07-31: ๅ็ๅ ฌ้๏ผใใซใกใใ็็จใฎใฟใปSHEL5K F1 0.837๏ผ
ใใผใฟใจๅธฐๅฑ
ๅญฆ็ฟใป่ฉไพกใฏไปฅไธใใผใฟใปใใใฎใขใใใผใทใงใณใใๅฐๅบใใฆใใพใ๏ผ่ฉไพกๅใฏๅญฆ็ฟใใ้คๅค๏ผ:
- SHEL5K๏ผCC BY 4.0, Otgonbold, M.E. et al.๏ผ
- Construction Site Safety๏ผCC BY 4.0, Roboflow Universe๏ผ
- Fall Detection๏ผCC BY 4.0, Roboflow Universe๏ผ
- FireExtinguisher๏ผCC BY 4.0, Roboflow Universe๏ผ
- Face Mask Detection๏ผCC0๏ผ
- Kitchen Hygiene Gear๏ผCC BY 4.0, Roboflow Universe๏ผ
- CAUCAFall๏ผCC BY 4.0, Universidad del Cauca โ Eraso Guerrero, J.C. et al., "Dataset for human fall recognition in an uncontrolled environment", Data in Brief, 2022๏ผ
- D-Fire๏ผde Venรขncio, P.V.A.B. et al., "An automatic fire detection system based on deep convolutional neural networks for low-power, resource-constrained devices", Neural Computing and Applications, 2022๏ผ
- Work at Height Safety Dataset๏ผCC BY 4.0, Universidad Carlos III de Madrid๏ผ
- PPEๆคๅบใใผใฟใปใใ๏ผCC BY 4.0๏ผ
- LOCO๏ผLogistics Objects in Context๏ผ๏ผCC0, Mayershofer, C. et al., "LOCO: Logistics Objects in Context", IEEE ICMLA 2020๏ผ
- MEVA๏ผMultiview Extended Video with Activities๏ผ๏ผCC BY 4.0, Corona, K. et al., WACV 2021 โ ่ปขๅ่ณชๅใฎ้ฐๆงใใฌใผใ ใซไฝฟ็จ๏ผ
- ๆฑๅ่ฉไพก: SHWD๏ผSafety Helmet Wearing Dataset๏ผ๏ผMIT๏ผ
- ใใผใน:
Qwen/Qwen3-VL-2B-Instruct๏ผApache-2.0๏ผ
Developed by tokimoa โ ใใผใฟใๅคใซๅบใใชใAIๆดป็จใๆฏๆดใใฆใใพใ
- Downloads last month
- 22
4-bit
Model tree for tokimoa/shokuba-probe-2b
Base model
Qwen/Qwen3-VL-2B-Instruct