QEV-9B / README.ko.md
querypieai's picture
Display QueryPie research license as plain text in all model cards
ed14203 verified
|
Raw History Blame Contribute Delete
10.8 kB
# QEV-9B
์ถ”๋ก  ์ „์šฉ ๋ฐฐํฌ ยท LoRA ๋ณ‘ํ•ฉ ยท BF16
**License: QueryPie AI Research License v1.0**
๊ฐœ์ธ์˜ ๋น„์ƒ์—…์  ์—ฐ๊ตฌยท์‹คํ—˜ ๋ฐ ๋น„์ƒ์—…์  ํ•™์ˆ  ์—ฐ๊ตฌยทํ‰๊ฐ€๋ฅผ ํ—ˆ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ƒ์—…์  ์ด์šฉ์—๋Š” QueryPie์˜ ์‚ฌ์ „ ์„œ๋ฉด ํ—ˆ๊ฐ€ ๋˜๋Š” ๋ณ„๋„์˜ ์ƒ์—…์šฉ ๋ผ์ด์„ ์Šค ๊ณ„์•ฝ์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ์ƒ์—…์  ์ด์šฉ์„ ํฌ๋งํ•˜๋Š” ๊ฒฝ์šฐ [QueryPie์— ๋ฌธ์˜ํ•ด ์ฃผ์„ธ์š”](https://www.querypie.com/ko/contact-us). ๊ธฐ์—… ์—…๋ฌดยท์ œํ’ˆ ๊ฐœ๋ฐœยท์ƒ์—… ๋ชฉ์  PoC๋Š” ๋ฌด๋ฃŒ ์—ฐ๊ตฌ ํ—ˆ์šฉ ๋ฒ”์œ„์— ํฌํ•จ๋˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ํฌํ•จ๋œ ์ œ3์ž ๊ธฐ๋ฐ˜ ๋ชจ๋ธ์—๋Š” ๊ฐ ์›์ฒœ ๋ผ์ด์„ ์Šค๊ฐ€ ์ ์šฉ๋ฉ๋‹ˆ๋‹ค.
[English](README.md) ยท [ํ•œ๊ตญ์–ด](README.ko.md) ยท [ๆ—ฅๆœฌ่ชž](README.ja.md)
## ๋ชจ๋ธ ์†Œ๊ฐœ
QEV-9B๋Š” Qwen3.5 ๊ธฐ๋ฐ˜์˜ ํ…์ŠคํŠธ ์˜์‚ฌ๊ฒฐ์ • ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. ๋ณ‘ํ•ฉ๋œ rank-16 LoRA์™€ pointer head๋ฅผ ์‚ฌ์šฉํ•˜๋ฉฐ choiceยทnoulยท์ˆœ์„œํ˜• score ์งˆ๋ฌธ์— ํ…์ŠคํŠธ ์ƒ์„ฑ ์—†์ด ์„ ํƒ์ง€ ํ™•๋ฅ ์„ ๋ฐ˜ํ™˜ํ•ฉ๋‹ˆ๋‹ค.
Base: `Qwen/Qwen3.5-9B-Base` ยท revision: `68c46c4b3498877f3ef123c856ecfde50c39f404` ยท format: `BF16`
## ํ•™์Šต๋ฐ์ดํ„ฐ
ํ”„๋กœ์ ํŠธ ํ•™์Šต๋ฐ์ดํ„ฐ๋Š” Public ๊ณต๊ฐœ๋ฐ์ดํ„ฐ์™€ ์‚ฌ๋‚ด DLP ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜๊ณ  ๋ฐ์ดํ„ฐ ์ฆ๊ฐ•์„ ์ ์šฉํ•ด ๊ตฌ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. ์‹ค์ œ ํ•™์Šตยท๊ฒ€์ฆยทcalibrationยทํ…Œ์ŠคํŠธ ๋ฐ์ดํ„ฐ๋Š” ๋ฐฐํฌํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ์ด ์ฒดํฌํฌ์ธํŠธ๋Š” ๊ณต๊ฐœ ๋ฐ์ดํ„ฐ ๊ธฐ๋ฐ˜ ๋ฒ”์šฉ ํ•™์Šต ํ›„ ์‚ฌ๋‚ด DLP ๋ฐ์ดํ„ฐ๋ฅผ ํ™œ์šฉํ•œ ๋„๋ฉ”์ธ ์ถ”๊ฐ€ ํ•™์Šต์„ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค.
## ๋…๋ฆฝํ˜• ํŒจํ‚ค์ง€
์ „์ฒด LoRA ๋ณ‘ํ•ฉ backbone, FP32 pointer head, tokenizer, calibration temperature์™€ Transformers ์ถ”๋ก  ์ฝ”๋“œ๋ฅผ ํฌํ•จํ•˜๋Š” ์ถ”๋ก  ์ „์šฉ ๋ฐฐํฌ์ž…๋‹ˆ๋‹ค. ๋ณ„๋„ adapter๋‚˜ ์™ธ๋ถ€ ๊ธฐ๋ฐ˜ ๋ชจ๋ธ ๋‹ค์šด๋กœ๋“œ๊ฐ€ ํ•„์š”ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ํ•™์Šต ์Šคํฌ๋ฆฝํŠธยทadapter ํŒŒ์ผยทoptimizer ์ƒํƒœยท์–‘์žํ™” ๋„๊ตฌยท๋ฐ์ดํ„ฐ์…‹์€ ๋ฐฐํฌํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ์ „์ฒด ๊ฐ€์ค‘์น˜ ๊ณต๊ฐœ ์ž์ฒด๊ฐ€ ์ถ”๊ฐ€ ํ•™์Šต์„ ๊ธฐ์ˆ ์ ์œผ๋กœ ์ฐจ๋‹จํ•˜์ง€๋Š” ์•Š์œผ๋ฉฐ, ์‚ฌ์šฉ์—๋Š” ํ•ด๋‹น ๋ผ์ด์„ ์Šค ์กฐ๊ฑด์ด ์ ์šฉ๋ฉ๋‹ˆ๋‹ค.
โ€˜์ถ”๋ก  ์ „์šฉโ€™์€ ์ œ๊ณต ํŒŒ์ผ์˜ ๊ตฌ์„ฑ์„ ๋œปํ•˜๋ฉฐ ์ถ”๊ฐ€ ํ•™์Šต์„ ๊ธฐ์ˆ ์ ์œผ๋กœ ์ฐจ๋‹จํ•œ๋‹ค๋Š” ์˜๋ฏธ๊ฐ€ ์•„๋‹™๋‹ˆ๋‹ค. LICENSE๊ฐ€ ํ—ˆ์šฉํ•˜๋Š” ๊ฐœ์ธ ๋น„์ƒ์—… ์—ฐ๊ตฌยท์‹คํ—˜๊ณผ ๋น„์ƒ์—… ํ•™์ˆ  ์—ฐ๊ตฌยทํ‰๊ฐ€ ๋ชฉ์ ์˜ ์ˆ˜์ •ยทํŒŒ์ธํŠœ๋‹์€ ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค. ์ƒ์—…ยท๊ธฐ์—… ์—…๋ฌด ๋ชฉ์ ์˜ ์‚ฌ์šฉ๊ณผ ๊ทธ ๋ชฉ์ ์˜ ํ‰๊ฐ€ยทํŒŒ์ธํŠœ๋‹์—๋Š” ์‚ฌ์ „ ์„œ๋ฉด ํ—ˆ๊ฐ€ ๋˜๋Š” ๋ณ„๋„ ์ƒ์—… ๊ณ„์•ฝ์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค.
LoRA๋ฅผ ๋ณ‘ํ•ฉํ•œ BF16 backbone๊ณผ ๋ณ€๊ฒฝํ•˜์ง€ ์•Š์€ FP32 pointer head๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋ณ‘ํ•ฉ์— ๋”ฐ๋ฅธ ๋ฐ˜์˜ฌ๋ฆผ ์ฐจ์ด๋Š” validation.json์— ๊ธฐ๋กํ•ฉ๋‹ˆ๋‹ค.
## ์‚ฌ์šฉ ์˜ˆ์‹œ
```bash
cd QEV-9B
pip install -r requirements.txt
python example.py --device cuda:0
```
```python
from transformers import AutoModel
model = AutoModel.from_pretrained(".", trust_remote_code=True,
device="cuda:0", local_files_only=True)
result = model.decide("This is a public product announcement.", {
"dlp": {"type": "noul", "instructions": "Does this contain personal contact information?",
"criteria": {"false": "No", "true": "Yes"}}
})
print(result)
```
์ „์ฒด ์ €์žฅ์†Œ๋ฅผ ๋‹ค์šด๋กœ๋“œํ•œ ๋’ค ์œ„ ๋ช…๋ น์„ ์‹คํ–‰ํ•˜์„ธ์š”. ๊ณต๊ฐœ๋œ Hub ๋ชจ๋ธ์„ ์ง์ ‘ ๋กœ๋“œํ•  ๋•Œ๋Š” "."์„ ํ•ด๋‹น ์ €์žฅ์†Œ ID๋กœ ๋ฐ”๊พธ๊ณ  local_files_only=True๋ฅผ ์ƒ๋žตํ•ฉ๋‹ˆ๋‹ค. AutoModel๊ณผ decide()๋ฅผ ์‚ฌ์šฉํ•˜๋ฉฐ, ์ผ๋ฐ˜ Transformers pipelineยทHub ์œ„์ ฏยทํ˜ธ์ŠคํŒ… Inference Providers์™€์˜ ํ˜ธํ™˜์„ฑ์€ ๊ฒ€์ฆํ•˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค.
## ๋ณ‘ํ•ฉ ์ „ BF16 ์›๋ณธ์˜ ์ฐธ๊ณ  ์„ฑ๋Šฅ
์•„๋ž˜ ์„ฑ๋Šฅํ‘œ๋Š” ๋ณ‘ํ•ฉ ์ „ BF16 ์›๋ณธ ์ฒดํฌํฌ์ธํŠธ์˜ ์—ญ์‚ฌ์  ํ‰๊ฐ€ ๊ฒฐ๊ณผ์ด๋ฉฐ, ์ด๋ฒˆ ๋ณ‘ํ•ฉยทINT8 ๊ณต๊ฐœ ํŒจํ‚ค์ง€์˜ ์‹ ๊ทœ ์‹ค์ธก ์„ฑ๋Šฅ์ด ์•„๋‹™๋‹ˆ๋‹ค. ์ „์ฒด ๋ฒค์น˜๋งˆํฌ๋Š” ์žฌ์‹คํ–‰ํ•˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ๋ณ‘ํ•ฉยท์–‘์žํ™”๋กœ ํ™•๋ฅ ์ด๋‚˜ ์„ ํƒ ๊ฒฐ๊ณผ๊ฐ€ ๋ฐ”๋€” ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. validation.json์€ ํ•œยท์˜ยท์ผ ํ•ฉ์„ฑ 9๋ฌธํ•ญ๊ณผ ์ €์žฅ ํ›„ ์žฌ๋กœ๋”ฉ ๊ฒ€์‚ฌ ๊ฒฐ๊ณผ๋งŒ ๊ธฐ๋กํ•ฉ๋‹ˆ๋‹ค.
| Metric | Value |
|---|---|
| MMLU-ProX-Lite ยท 16,384 questions | 44.51% |
| Belebele ยท 24,622 questions | 73.28% |
| Typed-Decision ยท 540 questions / 270 states | 87.96% |
| Equal-panel mean accuracy | 68.58% |
| DLP accuracy ยท 3,167 documents | 98.52% |
| DLP recall ยท threshold 0.50 | 97.81% |
| DLP F2 ยท threshold 0.50 | 97.95% |
DLP FP/FN: 19/28. Stored checkpoint temperature: 1.255943.
๋ฐ˜๋ณต ์‚ฌ์šฉ๋œ ์—ญ์‚ฌ์  ํ‰๊ฐ€์ด๋ฉฐ ์ƒˆ๋กœ์šด ๋…๋ฆฝ ํ™•์ธ ํ‰๊ฐ€๊ฐ€ ์•„๋‹™๋‹ˆ๋‹ค. ๊ณต๊ฐœ ๋ฒค์น˜๋งˆํฌ์˜ ์‚ฌ์ „ํ•™์Šต ์˜ค์—ผ ๊ฐ€๋Šฅ์„ฑ์„ ๋ฐฐ์ œํ•  ์ˆ˜ ์—†๊ณ  KOยทJA ๋ฒˆ์—ญ์€ ์›์–ด๋ฏผ ๊ฒ€์ˆ˜๋ฅผ ํ•˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ํ™•๋ฅ ๊ฐ’์€ ์ •๋‹ต๋ฅ  ๋ณด์ฆ์ด ์•„๋‹™๋‹ˆ๋‹ค.
## ํ‰๊ฐ€ ๋ฐ์ดํ„ฐ ์ถœ์ฒ˜์™€ QEV ์‹คํ—˜์šฉ ๊ฐ€๊ณต
์•„๋ž˜ ์ˆ˜์น˜๋Š” ์›์ฒœ ๋ฒค์น˜๋งˆํฌ ์ „์ฒด๋ฅผ ๊ทธ๋Œ€๋กœ ์‹คํ–‰ํ•œ ๊ณต์‹ ์ ์ˆ˜๊ฐ€ ์•„๋‹ˆ๋ผ, ์›์ฒœ ๋ฐ์ดํ„ฐ๋ฅผ QEV/KEV/JEV๊ฐ€ ๊ณตํ†ต์œผ๋กœ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ๋Š” typed-decision ์ž…๋ ฅ์œผ๋กœ ๊ฐ€๊ณตํ•œ ์‚ฌ๋‚ด ๋น„๊ต ํ‰๊ฐ€ ๊ฒฐ๊ณผ์ž…๋‹ˆ๋‹ค. ํ•™์Šต๋ฐ์ดํ„ฐ ๊ตฌ์„ฑ ์„ค๋ช…๊ณผ ํ‰๊ฐ€์…‹ ์ถœ์ฒ˜๋Š” ๊ตฌ๋ถ„ํ•ฉ๋‹ˆ๋‹ค. ์‹ค์ œ ์›์ฒœยท๊ฐ€๊ณต ๋ฐ์ดํ„ฐ๋Š” ์ด ์ €์žฅ์†Œ์— ํฌํ•จํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
- **MMLU-ProX-Lite**: [li-lab/MMLU-ProX-Lite](https://huggingface.co/datasets/li-lab/MMLU-ProX-Lite), MIT. ์งˆ๋ฌธ์„ state, ์ •๋‹ต ์„ ํƒ์„ choice๋กœ ๋ณ€ํ™˜ํ•˜๊ณ  cot_content ํ•ด์„ค์„ ์ž…๋ ฅ์—์„œ ์ œ์™ธํ–ˆ์œผ๋ฉฐ ๊ณต๋ž€ยทN/A ์„ ํƒ์ง€๋ฅผ ์ œ์™ธํ–ˆ์Šต๋‹ˆ๋‹ค. ์›๋ฌธ ๋ฌธ์ œยท์ •๋‹ต๊ณผ ์„ ํƒ์ง€ ์ˆœ์„œ๋ฅผ ์œ ์ง€ํ•˜๊ณ  ์œ ํšจ์„ฑ์„ ํ™•์ธํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ณตํ†ต 29๊ฐœ ์–ธ์–ด์˜ ์›์ฒœ 17,052๋ฌธํ•ญ์—์„œ ๊ธธ์ด ์ œํ•œ์„ ํ†ต๊ณผํ•œ ๊ณ ์ • ๋น„๊ต ๊ต์ง‘ํ•ฉ **16,384๋ฌธํ•ญ**์„ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค.
- **Belebele**: [facebook/belebele](https://huggingface.co/datasets/facebook/belebele), CC BY-SA 4.0. flores_passage๋ฅผ state, ์›๋ž˜ ์งˆ๋ฌธ์„ instructions, ๋„ค ์„ ํƒ์ง€๋ฅผ choice criteria๋กœ ๋ณ€ํ™˜ํ•˜๊ณ  correct_answer_num์„ ์ •๋‹ต ์„ ํƒ์ง€์— ์—ฐ๊ฒฐํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ณตํ†ต 29๊ฐœ ์–ธ์–ด์˜ ์›์ฒœ 26,100๋ฌธํ•ญ ์ค‘ ๊ฐ™์€ ๊ณ ์ • ๋น„๊ต ์กฐ๊ฑด์˜ **24,622๋ฌธํ•ญ**์„ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ฐ€๊ณต๋ณธ์€ ์‚ฌ๋‚ด ํ‰๊ฐ€์šฉ์ด๋ฉฐ ์›๋ฌธยท๊ฐ€๊ณต๋ณธ์„ ์™ธ๋ถ€ ์žฌ๋ฐฐํฌํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
- **๋น„๊ณต๊ฐœ Typed-Decision**: QueryPie ํ”„๋กœ์ ํŠธ์˜ private-general-v1. ์ •์ฑ…ยท๊ทœ์น™, ์ˆซ์žยท๋‚ ์งœ, ๋„๊ตฌ ๋ผ์šฐํŒ…, ์ •๋ณด ๋ถ€์กฑ, ์ˆœ์„œยท๋“ฑ๊ธ‰, ์ผ์ƒ ์ถ”๋ก ์„ ํ…œํ”Œ๋ฆฟ์œผ๋กœ ์ƒ์„ฑํ•˜๊ณ  ์ฝ”๋“œยท๊ทœ์น™ solver๋กœ ์ •๋‹ต์„ ๋งŒ๋“ค์—ˆ์Šต๋‹ˆ๋‹ค. ์˜์–ด๋ฅผ ํ•œ๊ตญ์–ดยท์ผ๋ณธ์–ด ํ…œํ”Œ๋ฆฟ์œผ๋กœ ๋ฒˆ์—ญํ•˜๋ฉด์„œ ์ •๋‹ตยท์„ ํƒ์ง€ keyยท์ˆซ์ž๋ฅผ ์œ ์ง€ํ•˜๊ณ  ์ž๋™ ์ •ํ•ฉ์„ฑ์„ ๊ฒ€์‚ฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์›์–ด๋ฏผ ๊ฒ€์ˆ˜๋Š” ์ˆ˜ํ–‰ํ•˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ์›์ฒœ 450๊ฐœ ์ƒํƒœยท900๋ฌธํ•ญ์—์„œ q1/q2 ์œ ํ˜•ยท์งˆ๋ฌธยท์„ ํƒ์ง€๊ฐ€ ๋™์ผํ•œ ์ƒํƒœ ์ „์ฒด๋ฅผ ์ œ์™ธํ•ด **270๊ฐœ ์ƒํƒœยท540๋ฌธํ•ญ**(์–ธ์–ด๋ณ„ 180๋ฌธํ•ญ)์„ ํ‰๊ฐ€ํ–ˆ์Šต๋‹ˆ๋‹ค.
- **DLP**: QueryPie ์‚ฌ๋‚ด ๋‹ค๊ตญ์–ด DLP ํ…Œ์ŠคํŠธ์…‹ **3,167๊ฑด**(EN 1,065 / JA 1,038 / KO 1,064). ์›๋ฌธยทIDยท์ •๋‹ต์„ ์œ ์ง€ํ•˜๊ณ  ๋ฌธ์„œ๋ฅผ state๋กœ, ๋ฏผ๊ฐ์ •๋ณด ์œ ๋ฌด๋ฅผ noul true/false ์งˆ๋ฌธ์œผ๋กœ ๋ณ€ํ™˜ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ •๋‹ต findings๊ฐ€ ํ•˜๋‚˜ ์ด์ƒ์ด๋ฉด ์–‘์„ฑ์œผ๋กœ ์ •์˜ํ•˜๋ฉฐ ์œ„์น˜ยท๊ฐœ์ˆ˜ยท์œ ํ˜•์˜ ์™„์ „ ์ถ”์ถœ์€ ํ‰๊ฐ€ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ์–‘์„ฑ 1,281 / ์ •์ƒ 1,886๊ฑด์ด๋ฉฐ threshold **0.50**์„ ์ ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ์‚ฌ๋‚ด ๋ฌธ์„œ ์›๋ฌธ๊ณผ ๋ผ๋ฒจ์€ ๋น„๊ณต๊ฐœ์ž…๋‹ˆ๋‹ค.
MMLUยทBelebele์€ state 384 / ์งˆ๋ฌธ๋ณ„ row 1,024 / packed 2,048์˜ ๊ธธ์ด ์กฐ๊ฑด์„ ํ†ต๊ณผํ•œ ๊ณ ์ • item-ID ๊ต์ง‘ํ•ฉ์œผ๋กœ ๋น„๊ตํ•˜๊ณ  ์ž๋™ ์ ˆ๋‹จํ•˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ๋ฒ”์šฉ ์ข…ํ•ฉ์€ MMLUยทBelebeleยทTyped-Decision์˜ ๋ฌธํ•ญ ๋‹จ์œ„ ์ •ํ™•๋„๋ฅผ ๋™์ผ ๊ฐ€์ค‘ ํ‰๊ท ํ•ฉ๋‹ˆ๋‹ค. ๊ณต๊ฐœ ๋ฐ์ดํ„ฐ์˜ ๊ธฐ๋ก๋œ revision๊ณผ ๊ฐ€๊ณต๋ณธ SHA-256์€ [EVALUATION_DATA_SOURCES.json](EVALUATION_DATA_SOURCES.json)์— ๋ณด์กดํ•˜๋ฉฐ ์›์ฒœ ๋ฐ์ดํ„ฐ์˜ ๋ผ์ด์„ ์Šค๋Š” QueryPie ๋ชจ๋ธ ๋ผ์ด์„ ์Šค๋กœ ๋ณ€๊ฒฝ๋˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
## ์ž…๋ ฅ ์ƒํ•œ๊ณผ ์‚ฌ์šฉ ๋ฒ”์œ„
๊ธฐ๋ณธ state ๊ฒ€์‚ฌ๋Š” len(encoded_state) + 1 โ‰ค 2,048์ด๋ฏ€๋กœ ์‚ฌ์šฉ์ž ๋ณธ๋ฌธ ์ž์ฒด๋Š” ํ•ด๋‹น tokenizer ๊ธฐ์ค€ ์ตœ๋Œ€ 2,047ํ† ํฐ์ž…๋‹ˆ๋‹ค. ๋ณธ๋ฌธยท์งˆ๋ฌธยท์„ ํƒ์ง€ยท๊ตฌ์กฐ/ํ…œํ”Œ๋ฆฟ ํ† ํฐ์„ ํฌํ•จํ•œ ์งˆ๋ฌธ๋ณ„ ์ „์ฒด row๋Š” ์ตœ๋Œ€ 2,688ํ† ํฐ์ด๋ฉฐ, ๋ชจ๋“  row ๊ธธ์ด์˜ ํ•ฉ๊ณ„๋Š” ์ตœ๋Œ€ 10,752ํ† ํฐ์ž…๋‹ˆ๋‹ค. ๋ฐ˜๋ณต๋˜๋Š” ๋ณธ๋ฌธ๋„ row๋งˆ๋‹ค ํ•ฉ์‚ฐํ•ฉ๋‹ˆ๋‹ค. ์งˆ๋ฌธ 1~4๊ฐœ, ์งˆ๋ฌธ๋ณ„ ์„ ํƒ์ง€ 2~10๊ฐœ๋ฅผ ์ง€์›ํ•˜๋ฉฐ ์ดˆ๊ณผ ์ž…๋ ฅ์€ ์ ˆ๋‹จ ์—†์ด ๊ฑฐ๋ถ€ํ•ฉ๋‹ˆ๋‹ค. ํ† ํฐ ์ˆ˜๋Š” ๋ชจ๋ธ๋ณ„ tokenizer์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์ง‘๋‹ˆ๋‹ค. ๊ธฐ์กด ํ’ˆ์งˆ ํ‰๊ฐ€๋Š” state 384 / row 1,024 / packed 2,048 ์กฐ๊ฑด์ด๋ฉฐ, ํ˜„์žฌ์˜ ๋” ๋†’์€ ์ž…๋ ฅ ์ƒํ•œ์—์„œ์˜ ํ’ˆ์งˆ์„ ์ž…์ฆํ•œ ๊ฒฐ๊ณผ๊ฐ€ ์•„๋‹™๋‹ˆ๋‹ค.
๋ฒ”์šฉ typed-decision๊ณผ ๋ฌธ์„œ ๋‹จ์œ„ DLP ๋ถ„๋ฅ˜๋ฅผ ์ง€์›ํ•ฉ๋‹ˆ๋‹ค. DLP์—์„œ๋Š” ๋ฏผ๊ฐ์ •๋ณด ์œ ๋ฌด๋ฅผ ํŒ๋‹จํ•˜๋ฉฐ finding ์œ„์น˜ยท๊ฐœ์ˆ˜ยท์œ ํ˜•์„ ์ถ”์ถœํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. choice๋Š” ์ด๋ฆ„ ์žˆ๋Š” ์„ ํƒ์ง€, noul์€ false/true criteria์™€ true ํ™•๋ฅ , score๋Š” ์ˆœ์„œ๊ฐ€ ์žˆ๋Š” ๋ชฉ๋ก์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. score๋Š” ์—ฐ์†ํ˜• ์ ์ˆ˜ ๋Œ€์‹  0๋ถ€ํ„ฐ ์‹œ์ž‘ํ•˜๋Š” ์„ ํƒ ์ธ๋ฑ์Šค์™€ ๊ฐ ์ธ๋ฑ์Šค์˜ ํ™•๋ฅ ์„ ๋ฐ˜ํ™˜ํ•ฉ๋‹ˆ๋‹ค. confidence๋Š” ์ตœ๋Œ€ ์„ ํƒ์ง€ ํ™•๋ฅ ์ด๋ฉฐ ์‹ค์ธก ์ •๋‹ต๋ฅ ์ด ์•„๋‹™๋‹ˆ๋‹ค. ํ…์ŠคํŠธ๋ฅผ ์ƒ์„ฑํ•˜์ง€ ์•Š์œผ๋ฉฐ ์ž„์˜ ๋„๋ฉ”์ธ์—์„œ์˜ ์‹ ๋ขฐ๋„๋ฅผ ๋ณด์žฅํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
๊ฒ€์ฆ ์žฅ๋น„๋Š” NVIDIA H200/CUDA์ž…๋‹ˆ๋‹ค. CPUยท๋‹ค๋ฅธ ๋Ÿฐํƒ€์ž„์˜ ์„ฑ๋Šฅ์€ ๊ฒ€์ฆํ•˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ์ „์šฉ ์ฝ”๋“œ ์‹คํ–‰์—๋Š” trust_remote_code=True๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค.
## ๋ผ์ด์„ ์Šค์™€ ์›์ฒœ ํ‘œ์‹œ
QueryPie์˜ ํ•™์Šต๋œ ์ˆ˜์ •๋ถ„(๋ณ‘ํ•ฉ LoRA ํฌํ•จ)ยทpointer headยท์ „์šฉ ์ฝ”๋“œ์—๋Š” QueryPie AI Research License v1.0์„ ์ ์šฉํ•ฉ๋‹ˆ๋‹ค. ์›๋ž˜ Qwen/Gemma ๊ธฐ๋ฐ˜ ๊ฐ€์ค‘์น˜ยทtokenizer์—๋Š” Apache-2.0์„ ์œ ์ง€ํ•˜๋ฉฐ, ๊ทธ ์›์ฒœ ๋ผ์ด์„ ์Šค๋กœ ์ด๋ฏธ ๋ถ€์—ฌ๋œ ๊ถŒ๋ฆฌ๋Š” ์ œํ•œํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ๊ตฌ์„ฑ ์š”์†Œ๋ณ„ ๋ฒ”์œ„๋Š” [THIRD_PARTY_LICENSES.md](THIRD_PARTY_LICENSES.md), ์›๋ฌธ์€ [LICENSES/Apache-2.0.txt](LICENSES/Apache-2.0.txt), ์›์ฒœ ํ‘œ์‹œ๋Š” NOTICE๋ฅผ ์ฐธ์กฐํ•˜์„ธ์š”. ํ•™์Šต๋ฐ์ดํ„ฐ ์ž์ฒด๋ฅผ ์žฌ๋ผ์ด์„ ์Šคํ•˜๊ฑฐ๋‚˜ ๋ฐฐํฌํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
๊ฐœ์ธ์˜ ๋น„์ƒ์—…์  ์—ฐ๊ตฌยท์‹คํ—˜ ๋ฐ ๋น„์ƒ์—…์  ํ•™์ˆ  ์—ฐ๊ตฌยทํ‰๊ฐ€๋ฅผ ํ—ˆ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ƒ์—…์  ์ด์šฉ์—๋Š” QueryPie์˜ ์‚ฌ์ „ ์„œ๋ฉด ํ—ˆ๊ฐ€ ๋˜๋Š” ๋ณ„๋„์˜ ์ƒ์—…์šฉ ๋ผ์ด์„ ์Šค ๊ณ„์•ฝ์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ์ƒ์—…์  ์ด์šฉ์„ ํฌ๋งํ•˜๋Š” ๊ฒฝ์šฐ [QueryPie์— ๋ฌธ์˜ํ•ด ์ฃผ์„ธ์š”](https://www.querypie.com/ko/contact-us). ๊ธฐ์—… ์—…๋ฌดยท์ œํ’ˆ ๊ฐœ๋ฐœยท์ƒ์—… ๋ชฉ์  PoC๋Š” ๋ฌด๋ฃŒ ์—ฐ๊ตฌ ํ—ˆ์šฉ ๋ฒ”์œ„์— ํฌํ•จ๋˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ํฌํ•จ๋œ ์ œ3์ž ๊ธฐ๋ฐ˜ ๋ชจ๋ธ์—๋Š” ๊ฐ ์›์ฒœ ๋ผ์ด์„ ์Šค๊ฐ€ ์ ์šฉ๋ฉ๋‹ˆ๋‹ค.
๋‚ด๋ถ€ ์ž์‚ฐ ๋ˆ„๋ฝ์„ ๋ง‰๊ธฐ ์œ„ํ•ด generic save_pretrained()๋Š” ๋น„ํ™œ์„ฑํ™”ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ €์žฅ์†Œ ์ „์ฒด๋ฅผ ๋ณต์‚ฌํ•˜๊ฑฐ๋‚˜ ๋‹ค์šด๋กœ๋“œํ•˜์„ธ์š”. ํŒจํ‚ค์ง€ ๊ฒ€์ฆ์€ validation.json๊ณผ MANIFEST.sha256์— ๊ธฐ๋ก๋ฉ๋‹ˆ๋‹ค.
## ๋ฐฐํฌ ๊ฒ€์ฆ
| ๋น„๊ต | ์ตœ๋Œ€ ํ™•๋ฅ  ์ฐจ์ด | ์„ ํƒ ๊ฒฐ๊ณผ ๋ณ€ํ™” / 9๋ฌธํ•ญ |
|---|---:|---:|
| ์›๋ณธ BF16 โ†’ ๋ณ‘ํ•ฉ BF16 | 0.00107431 | 0 |
| ์ €์žฅ๋œ ๋ชจ๋ธ โ†’ ์žฌ๋กœ๋”ฉ ๋ชจ๋ธ | 0.00000000 | 0 |
[validation.json](validation.json) ยท [MANIFEST.sha256](MANIFEST.sha256)