How to use from
llama.cpp
Install (macOS, Linux)
curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf interfaceinde/qwen35-4b-interface-classifier-GGUF:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf interfaceinde/qwen35-4b-interface-classifier-GGUF:Q4_K_M
Install from WinGet (Windows)
winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf interfaceinde/qwen35-4b-interface-classifier-GGUF:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf interfaceinde/qwen35-4b-interface-classifier-GGUF:Q4_K_M
Build from source code
git clone https://github.com/ggml-org/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-app
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama serve -hf interfaceinde/qwen35-4b-interface-classifier-GGUF:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama cli -hf interfaceinde/qwen35-4b-interface-classifier-GGUF:Q4_K_M
Use Docker
docker model run hf.co/interfaceinde/qwen35-4b-interface-classifier-GGUF:Q4_K_M
Quick Links

INTERFACE Qwen3.5 4B ๋ฉ”์ผ ๋ถ„๋ฅ˜ ๋ชจ๋ธ

INTERFACE ํ”„๋กœ์ ํŠธ์˜ ํ•œ๊ตญ์–ด ๋ฉ”์ผ 20ํ•ญ๋ชฉ JSON ๋ถ„๋ฅ˜์šฉ ๋ชจ๋ธ์ด๋‹ค. ์ „์ฒด ํ•™์Šต ์‹คํ–‰์—์„œ ๊ฐœ๋ฐœ ๊ฒ€์ฆ ํ•ญ๋ชฉ ์ผ์น˜์œจ์ด ๊ฐ€์žฅ ๋†’์•˜๋˜ checkpoint 500๋งŒ ๋ฐฐํฌํ•œ๋‹ค. ์ตœ์ข… ์ €์žฅ checkpoint 650๊ณผ ๋‹ค๋ฅธ ํ•™์Šต ์ฒดํฌํฌ์ธํŠธ๋Š” ํฌํ•จํ•˜์ง€ ์•Š๋Š”๋‹ค.

์ด ๋ชจ๋ธ์€ ์ผ๋ฐ˜ ๋Œ€ํ™”ยท๋ฉ”์ผ ์ดˆ์•ˆ ์ž‘์„ฑยทOCR ๋ชจ๋ธ์˜ ๋Œ€์ฒด๋ณธ์ด ์•„๋‹ˆ๋‹ค. ํ…์ŠคํŠธ ์ „์šฉ GGUF์ด๋ฉฐ ์ด๋ฏธ์ง€ projector๋ฅผ ํฌํ•จํ•˜์ง€ ์•Š๋Š”๋‹ค.

Ollama ์‚ฌ์šฉ

ollama run hf.co/interfaceinde/qwen35-4b-interface-classifier-GGUF:Q4_K_M

ํŒŒ์ผ ๋‹ค์šด๋กœ๋“œ ํ›„ ๋กœ์ปฌ ์ด๋ฆ„์œผ๋กœ ๋“ฑ๋กํ•  ์ˆ˜๋„ ์žˆ๋‹ค.

ollama create interface-qwen35-classifier:step500-q4_k_m -f Modelfile

์ž…๋ ฅ์—๋Š” classifier_instruction.txt์˜ ์ „์ฒด ์ง€์‹œ๋ฌธ์„ ๋ถ™์ด๊ณ  ๊ทธ ๋’ค์— ๋ฉ”์ผ์˜ input_text๋ฅผ ์ œ๊ณตํ•œ๋‹ค. ์‘๋‹ต ํ˜•์‹์€ response_schema.json์— ์ •์˜ํ•œ๋‹ค. API ํ˜ธ์ถœ์—์„œ๋Š” think: false, stream: false, temperature: 0, ์ถœ๋ ฅ ํ•œ๋„ 1024๋ฅผ ์‚ฌ์šฉํ•œ๋‹ค. 500ํ†ต ๊ฐœ๋ฐœ ๊ฒ€์ฆ๊ณผ ๋™์ผํ•œ ์ž…๋ ฅ ํฌ๋งท์€ ๋ณ„๋„๋กœ ์œ ์ง€ํ•ด์•ผ ํ•œ๋‹ค.

๋ฐฐํฌ ๋ฒ„์ „๊ณผ ์—…๋ฐ์ดํŠธ

  • BEST: ํ˜„์žฌ ๋ฐฐํฌํ•  ๋‹จ์ผ GGUF ํŒŒ์ผ๋ช….
  • manifest.json: ์›๋ณธ ๋ชจ๋ธยท์ฒดํฌํฌ์ธํŠธยท๋ณ€ํ™˜ ๋„๊ตฌยทGGUF SHA256.
  • ์ƒˆ ์—…๋กœ๋“œ ํ›„ ๊ธฐ์กด Ollama ๋ชจ๋ธ์ด ์ž๋™ ๊ต์ฒด๋˜์ง€๋Š” ์•Š๋Š”๋‹ค. ๋‹ค์‹œ ollama pullํ•˜๊ฑฐ๋‚˜ ๋ณ„๋„์˜ ๊ฐฑ์‹  ์ ˆ์ฐจ๊ฐ€ ํ•„์š”ํ•˜๋‹ค.
  • INTERFACE ์•ฑ์˜ ๊ธฐ์กด OLLAMA_MODEL๊ณผ ์ดˆ์•ˆ ์ƒ์„ฑ ๋ชจ๋ธ์€ ์ด๋ฒˆ ๊ณต๊ฐœ ์—…๋กœ๋“œ๋กœ ๋ณ€๊ฒฝํ•˜์ง€ ์•Š์•˜๋‹ค.

ํ•™์Šต ์ถœ์ฒ˜

  • ์›๋ณธ: Qwen/Qwen3.5-4B
  • ๊ณ ์ • ์›๋ณธ ๋ฆฌ๋น„์ „: 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a
  • QLoRA NF4, BF16 ๊ณ„์‚ฐ, LoRA rank 4 / alpha 8, ํ•™์Šต ํŒŒ๋ผ๋ฏธํ„ฐ 458,752๊ฐœ.
  • 1 epoch ๊ณ„ํš, microbatch 1, gradient accumulation 4, AdamW, ํ•™์Šต๋ฅ  0.0001, seed 42.
  • ํ•ฉ์„ฑ ๋ฉ”์ผ train ๋ถ„ํ•  4,000ํ†ต ์ค‘ ๊ฒ€ํ†  ํ›„๋ณด๋ฅผ ์ œ์™ธํ•œ 2,868ํ†ต ์‚ฌ์šฉ.
  • ์‹คํ–‰์€ 650๋‹จ๊ณ„์—์„œ 3ํšŒ ์—ฐ์† ๋ฏธ๊ฐœ์„ ์œผ๋กœ ์ข…๋ฃŒ. ๋ฐฐํฌ checkpoint 500์€ ๋ˆ„์  ํ•™์Šต 2,000ํ†ต ์ง€์ ์ด๋‹ค.
  • ์›๋ณธ BF16 ๊ฐ€์ค‘์น˜์— checkpoint 500 LoRA๋ฅผ ๋ณ‘ํ•ฉํ•˜๊ณ  llama.cpp b10969 ๋ณ€ํ™˜๊ธฐ๋กœ ํ…์ŠคํŠธ GGUF๋ฅผ ๋งŒ๋“ค์—ˆ๋‹ค.
  • ์„ค์น˜๋œ Ollama 0.34.3์ด ๊ณ ์ •ํ•œ llama.cpp ๋ฒ„์ „๊ณผ ๋งž์ถ˜ ๋ณ€ํ™˜๊ธฐ์ด๋ฉฐ, Q4_K_M ์–‘์žํ™” ๋„๊ตฌ๋Š” b12556์ด๋‹ค.

ํ•™์Šต ๋ฐ์ดํ„ฐยท๊ฒ€์ฆ ๋ฉ”์ผ ์›๋ฌธยท์ •๋‹ต ํŒŒ์ผยทoptimizer ์ƒํƒœยท๋กœ์ปฌ ๊ฒฝ๋กœยท์ธ์ฆ ์ •๋ณด๋Š” ์ด ์ €์žฅ์†Œ์— ์—…๋กœ๋“œํ•˜์ง€ ์•Š๋Š”๋‹ค.

์ถœ์ฒ˜ ์ฒดํฌํฌ์ธํŠธ์˜ ๊ฐœ๋ฐœ ๊ฒ€์ฆ ๊ฒฐ๊ณผ

๋‹ค์Œ ์ˆ˜์น˜๋Š” GGUF ๋ณ€ํ™˜ ์ „์˜ native NF4 QLoRA checkpoint 500์—์„œ ์ธก์ •ํ•˜์˜€๋‹ค. ๋ณ€ํ™˜ยท์žฌ์–‘์žํ™” ํ›„ ๋™์ผ ์„ฑ๋Šฅ์„ ๋ณด์žฅํ•˜์ง€ ์•Š์œผ๋ฉฐ, GGUF ์ „์ฒด 500ํ†ต ์žฌํ‰๊ฐ€์™€ ๋…๋ฆฝ ํ…Œ์ŠคํŠธ๋Š” ๋ฏธ์‹คํ–‰์ด๋‹ค.

Ollama 0.34.3์—์„œ ์ด GGUF์˜ ์‹ค์ œ ๋กœ๋“œ์™€ ๊ธฐ์กด ๊ฐœ๋ฐœ ๊ฒ€์ฆ 1๊ฑด ์‹คํ–‰์„ ํ™•์ธํ•˜์˜€๋‹ค. ํ•ด๋‹น 1๊ฑด์€ JSON ํŒŒ์‹ฑยท์Šคํ‚ค๋งˆ ๊ฒ€์‚ฌ๋ฅผ ํ†ต๊ณผํ•˜๊ณ  20ํ•ญ๋ชฉ ์ค‘ 19ํ•ญ๋ชฉ์ด ์ผ์น˜ํ•˜์˜€๋‹ค. ์ด๋Š” ๋กœ์ปฌ ์‹คํ–‰ยท์ถœ๋ ฅ ๋™์ž‘ ํ™•์ธ์ด๋ฉฐ 500ํ†ต ํ’ˆ์งˆ ํ‰๊ฐ€๋ฅผ ๋Œ€์ฒดํ•˜์ง€ ์•Š๋Š”๋‹ค.

์ง€ํ‘œ ๊ฒฐ๊ณผ
๊ฐœ๋ฐœ ๊ฒ€์ฆ ๊ทœ๋ชจ 500ํ†ต ร— 20ํ•ญ๋ชฉ
ํ•ญ๋ชฉ ์—„๊ฒฉ ์ผ์น˜ 9,288 / 10,000 ยท 92.88%
JSON ํŒŒ์‹ฑ ์„ฑ๊ณต 500 / 500 ยท 100%
์Šคํ‚ค๋งˆ ํ†ต๊ณผ 498 / 500 ยท 99.6%
์ „์ฒด 20ํ•ญ๋ชฉ ์ผ์น˜ ๋ฐ ์Šคํ‚ค๋งˆ ํ†ต๊ณผ 31 / 500 ยท 6.2%

ํ•ญ๋ชฉ์€ ๋™์ผ ๋น„์ค‘์œผ๋กœ ์—„๊ฒฉ ๋น„๊ตํ•œ๋‹ค. action_items๋Š” 37/500(7.4%), routing์€ 408/500(81.6%)์ด๋‹ค. ๋ฌธ์žฅ ํ‘œํ˜„ยท๋ฐฐ์—ด ์ˆœ์„œ ์ฐจ์ด๋„ ๋ถˆ์ผ์น˜๊ฐ€ ๋  ์ˆ˜ ์žˆ๋‹ค. ๊ฐ™์€ ๊ฐœ๋ฐœ ๊ฒ€์ฆ ์„ธํŠธ๋ฅผ ๋ฐ˜๋ณต ์‚ฌ์šฉํ–ˆ์œผ๋ฉฐ, ์ •๋‹ต์˜ ๋…๋ฆฝ์ ์ธ ์˜๋ฏธ ๊ฒ€์ˆ˜์™€ ์‹ค์ œ ์—…๋ฌด ์„ฑ๊ณต๋ฅ ์€ ์ด๋ฒˆ ๊ฒฐ๊ณผ๋กœ ํ™•์ธํ•˜์ง€ ์•Š์•˜๋‹ค. ์ถœ๋ ฅ ํ˜•์‹ ์„ฑ๊ณต์„ ๋‚ด์šฉ ์ •ํ™•๋„๋‚˜ ํ˜„์—… ์ž๋™ ์ฒ˜๋ฆฌ ๊ฐ€๋Šฅ์„ฑ์œผ๋กœ ํ•ด์„ํ•˜์ง€ ์•Š๋Š”๋‹ค.

ํ•™์Šต ๋‚ด์—ญ๊ณผ ์„ค๋ช… ํŒŒ์ผ

  • learning_history.md: ๋ฐ์ดํ„ฐ ๊ตฌ์„ฑ, ํ•™์Šต ์„ค์ •, ์‹œ์ž‘ยท์ข…๋ฃŒ, ์กฐ๊ธฐ ์ข…๋ฃŒ์™€ ์ตœ๊ณ  ์ฒดํฌํฌ์ธํŠธ ์„ ์ •, ์ „์ฒด 13ํšŒ ๊ฒ€์ฆ ๊ฒฐ๊ณผ.
  • training_history.csv: ๋ฉ”์ผ ์›๋ฌธ ์—†์ด 650๊ฐœ optimizer ๋‹จ๊ณ„์˜ ์†์‹คยท๊ธฐ์šธ๊ธฐยท์ฒ˜๋ฆฌ๋Ÿ‰ยท์‹œ๊ฐ„ ๊ธฐ๋ก.
  • validation_history.json: 50~650๋‹จ๊ณ„์˜ ๊ฒ€์ฆ ์ง‘๊ณ„.
  • classifier_instruction.txt: ํ•™์Šต ์‹œ ์‚ฌ์šฉํ•œ ๊ณตํ†ต ๋ถ„๋ฅ˜ ์ง€์‹œ๋ฌธ.
  • response_schema.json: ์ถœ๋ ฅ 20ํ•ญ๋ชฉ์˜ ๊ตฌ์กฐ์™€ ํ—ˆ์šฉ๊ฐ’.
  • manifest.json: ๋ฐฐํฌ ๊ฐ€์ค‘์น˜ ์ถœ์ฒ˜ยท๋„๊ตฌยทSHA256ยท๊ฒ€์ฆ ๋ฒ”์œ„.

๋ผ์ด์„ ์Šค

์›๋ณธ Qwen3.5-4B์˜ Apache-2.0 ๋ผ์ด์„ ์Šค์™€ ๊ณ ์ง€ ์กฐ๊ฑด์„ ๋”ฐ๋ฅธ๋‹ค. ์›๋ณธ LICENSE๋ฅผ ๋™๋ด‰ํ•˜์˜€๋‹ค.

Downloads last month
70
GGUF
Model size
4B params
Architecture
qwen35
Hardware compatibility
Log In to add your hardware

4-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for interfaceinde/qwen35-4b-interface-classifier-GGUF

Finetuned
Qwen/Qwen3.5-4B
Finetuned
(1,552)
this model