ko-anti-cliche-qwen3.6-27b
A LoRA adapter for Qwen/Qwen3.6-27B that suppresses the translationese "AI tells" Korean LLM output tends to carry โ ๋จ์ํ ~๊ฐ ์๋๋ผ ("not just X, but Y"), ํต์ฌ์ ~์ ์๋ค ("the key lies in X"), ๋ฐ๋ง ๋ง๋ค ("that's only half true"). Same recipe as the Gemma siblings ko-anti-cliche-gemma4-26b-a4b / -gemma4-12b.
What it suppresses
๋จ์ํ ~๊ฐ ์๋ B โ a manufactured contrast (English "not just X, but Y")
๋จ์ํ ํต์ ๊ฐ ์๋ ๋น์ ์ด๋ค โ ๋น์ ์ ์ ์ํ๋ ๋ฅ๋ ฅ์ด๋ค
Genuine factual distinctions like GPU ๋ฌธ์ ๊ฐ ์๋๋ผ ๋๋ผ์ด๋ฒ ๋ฌธ์ ๋ค are natural in Korean and are left untouched.
ํต์ฌ์ ~์ ์๋ค โ a "lies in" calque ยท ๋ฐ๋ง ๋ง๋ค โ an "only half true" calque
Training
| Base | Qwen/Qwen3.6-27B (dense) |
| Adapter | LoRA, r=16, ฮฑ=32, dropout=0.05, target q,k,v,o (language tower) |
| Objective | 1.0ยทlocalized-DPO + 0.3ยทclean-anchor-CE + 0.7ยทtoken-unlikelihood |
| Data | 428 preference pairs (nabb 311 / ultimately 85 / half_right 32) |
| Schedule | ฮฒ=0.3, lr 2e-5, 2 epochs |
Trained in non-thinking format (empty <think></think> block) so the style transfers to normal answers. Localized DPO on the scaffold span, token-unlikelihood on scaffold tokens (factual rows exempt), CE anchor for fluency.
Self-evaluation
Protocol. 1,375 synthetic definition prompts; the 1,375โbase-positive subset (base emitted not-A-but-B) evaluated. Base and adapter decoded through the same HF stack, enable_thinking=False (non-thinking, matching training). Truncated rows dropped โ N = 1,068. Regex proxy, 95% Wilson CIs.
Table 1 โ Conditional removal rate
| Target pattern | n | Removed | 95% CI |
|---|---|---|---|
โ ๋จ์ํ/๋จ์ํ forced contrast |
994 | 94.9% | 93.3 โ 96.1 |
โก ~์ ์๋ค lies-in calque |
4 | 100.0% | 51 โ 100 |
| not-A-but-B (any form) | 1,068 | 89.8% | 87.8 โ 91.5 |
Table 2 โ Auxiliary measures (N = 1,068)
| Measure | Base | Adapter |
|---|---|---|
๊ฒฐ๊ตญ frequency (token count) |
54 | 14 |
| Degenerate outputs | โ | 4 (0.4%) |
| Length ratio, adapter รท base (median) | โ | 0.61 |
| New scaffold family introduced (base-positive) | โ | 0 / 1,068 |
Note: on Qwen3.6-27B the adapter removes not-A-but-B in all forms most aggressively of the three siblings (89.8% vs 55.7% / 73.1%), and outputs are more concise (length 0.61).
Limitations
Surface-level (word/phrase) tool. English and Korean differ in paragraph/argument organization (point-first vs. point-last); a LoRA does not change that discourse structure. Evaluated in non-thinking mode; behavior inside <think> reasoning traces is out of scope.
Usage
from peft import PeftModel
from transformers import AutoModelForImageTextToText, AutoTokenizer
base = AutoModelForImageTextToText.from_pretrained(
"Qwen/Qwen3.6-27B", torch_dtype="bfloat16", device_map="auto")
tok = AutoTokenizer.from_pretrained("Qwen/Qwen3.6-27B")
model = PeftModel.from_pretrained(base, "sdgb-ai/ko-anti-cliche-qwen3.6-27b")
# generate with enable_thinking=False for normal (non-reasoning) answers
ํ๊ตญ์ด
๋ณธ LoRA๋ Qwen3.6๊ฐ ํ๊ตญ์ด๋ฅผ ์์ฑํ ๋ ๋ํ๋๋ "๋จ์ํ ~๊ฐ ์๋๋ผ", "ํต์ฌ์ ~์ ์๋ค", "๋ฐ๋ง ๋ง๋ค" ๊ฐ์ LLM ํน์ ์ ๋ฒ์ญํฌ๋ฅผ ์ต์ ํ๊ณ , ์์ฐ์ค๋ฌ์ด ํ๊ตญ์ด๋ก ๊ต์ ํ๋ค. Gemma ํ์ ๋ค๊ณผ ๋์ผ ๋ ์ํผ.
์ต์ ๋์
"๋จ์ํ ~๊ฐ ์๋ B" โ B๋ฅผ ๊ฐ์กฐํ๊ธฐ ์ํด ๊น์๋ด๋ฆด A๋ฅผ ์ธ์์ ์ผ๋ก ๋ง๋ค์ด ๋ด์ธ์ฐ๋ ์์ด์ ๊ตฌ๋ฌธ.
๋จ์ํ ํต์ ๊ฐ ์๋ ๋น์ ์ด๋ค โ ๋น์ ์ ์ ์ํ๋ ๋ฅ๋ ฅ์ด๋ค ๋ค๋ง "GPU ๋ฌธ์ ๊ฐ ์๋๋ผ ๋๋ผ์ด๋ฒ ๋ฌธ์ ๋ค"์ฒ๋ผ ์ค์ ์ฌ์ค์ ๊ตฌ๋ถํ๋ ํํ์ ์ต์ ํ์ง ์์.
"ํต์ฌ์ ~์ ์๋ค" โ ์์ด "the key lies in"์ ์ง์ญ ยท "๋ฐ๋ง ๋ง๋ค" โ "that's only half true"์ ์ง์ญ.
ํ์ต ๋ฐฉ๋ฒ
| ๋ฒ ์ด์ค | Qwen/Qwen3.6-27B (dense) |
| ์ด๋ํฐ | LoRA, r=16, ฮฑ=32, dropout=0.05, target q,k,v,o |
| ๋ชฉ์ ํจ์ | 1.0ยทlocalized-DPO + 0.3ยทclean-anchor-CE + 0.7ยทtoken-unlikelihood |
| ๋ฐ์ดํฐ | ์ ํธ์ 428๊ฐ (nabb 311 / ultimately 85 / half_right 32) |
| ์ค์ผ์ค | ฮฒ=0.3, lr 2e-5, 2 epoch |
thinking ๋ชจ๋ธ์ด๋ผ non-thinking ํฌ๋งท(๋น <think></think>)์ผ๋ก ํ์ตํด ์ผ๋ฐ ๋ต๋ณ์ ์คํ์ผ์ด ์ ์ด๋๊ฒ ํจ.
์์ฒด ํ๊ฐ
synthetic ์ ์ํ ํ๋กฌํํธ ์ค base-positive๋ฅผ base/adapter ๋์ผ HF ์คํ + enable_thinking=False ๋ก ์์ฑ, truncated ์ ์ธ โ N = 1,068. ์ ๊ท์ proxy, 95% Wilson CI.
| ํ์ ํจํด | n | ์ ๊ฑฐ | 95% CI |
|---|---|---|---|
โ ๋จ์ํ/๋จ์ํ ๊ฐ์ ๋์กฐ |
994 | 94.9% | 93.3 โ 96.1 |
โก ~์ ์๋ค lies-in |
4 | 100.0% | 51 โ 100 |
| not-A-but-B (์ ์ฒด) | 1,068 | 89.8% | 87.8 โ 91.5 |
๋ณด์กฐ: ๊ฒฐ๊ตญ 54โ14 ยท degenerate 4(0.4%) ยท ๊ธธ์ด๋น 0.61 ยท ์ family ๋์
0/1,068. ์ธ ํ์ ์ค not-A-but-B ์ ์ฒด ์ ๊ฑฐ๊ฐ ๊ฐ์ฅ ๊ฐํจ.
ํ๊ณ
ํ๋ฉด ์ดํยท๊ตฌ๋ฌธ ๋๊ตฌ. ์์ด/ํ๊ตญ์ด ๋ดํ ๊ตฌ์กฐ(๋๊ด์ vs ๋ฏธ๊ด์) ์ฐจ์ด๊น์ง๋ ๋ชป ๋ฐ๊ฟ. non-thinking ๋ชจ๋๋ก ํ๊ฐํ์ผ๋ฉฐ <think> ์ถ๋ก ํธ๋ ์ด์ค ๋ด๋ถ๋ ๋ฒ์ ๋ฐ.
LoRA r=16, q/k/v/o (language tower). base Qwen/Qwen3.6-27B (apache-2.0). ํ์ : ko-anti-cliche-gemma4-26b-a4b, ko-anti-cliche-gemma4-12b.