chipsnug/ko-quant-loss-v0

Chipsnug koqloss (v0.2.0)

This dataset measures how much more information Korean loses than English when an open model is quantized, using the same content in both languages. It is a mirror of the result tables in https://github.com/chipsnug/koqloss. The full report, in English and then Korean, is in koqloss-public.md.

• Q1 — KL vs Q8_0 on parallel text (FLORES-101 devtest, sentences 1–300). For the same content, Korean loses 1.09–3.45× more than English; the 95% interval is above 1.0 in 7 of 8 model × level combinations (not for A.X-4.0-Light at Q3_K_M, 0.97–1.23). Per token, kanana-1.5-2.1b-instruct shows no detectable extra loss (95% CIs 0.84–1.14), A.X-4.0-Light adds 1.17–1.32× (the hypothesis that Korean-focused models have no per-token gap is rejected), Qwen3 adds 1.42–2.04×. A.X-4.0-Light needs 0.93× as many tokens for Korean, so its user-facing gap is the smallest.
• Q2 — paired multiple choice (300 MMMLU KO_KR ↔ MMLU items). The KO−EN flip gap is −1.7 to +18.5 points (higher in Korean in 7 of 8 combinations). Intervals exclude 0 for Qwen3-1.7B at both levels and, only just, for A.X-4.0-Light at Q4_K_M.
• Q3 — speed and memory on an Apple M4 Pro (Metal). Q4_K_M generates fastest for all four models.

Models (GGUF, three levels from the same repo, no self-quantization):

• kakaocorp/kanana-1.5-2.1b-instruct-2505 (DevQuasar)
• skt/A.X-4.0-Light (mykor), added 2026-09-28
• Qwen/Qwen3-1.7B (bartowski)
• Qwen/Qwen3-4B (bartowski)

Q8_0 is the reference, not BF16. One device, one session, 300 sentences, 300 item pairs.

Files

File	Contents
data/kl.csv	Q1 ratios with 95% intervals
data/mc.csv	Q2 accuracy, flip rates and gap with 95% intervals
data/speed.csv	Q3 tokens/s and max RSS
data/models.csv	GGUF repos, files, sizes and SHA-256
data/inputs.json	Input IDs, build rules and SHA-256. No source text
data/run-*.json	Per-chunk KL, per-item correct bits, bench output
data/summary.json	All tables in one file

License

• Tables and documentation: CC BY 4.0.
• The FLORES-101-derived fields in data/inputs.json (q1_kl_document: sentence IDs, hashes, build rule) also follow CC BY-SA 4.0, the license of FLORES-101. See THIRD_PARTY_NOTICES.md.
• MMMLU and MMLU are MIT. Only row positions and hashes are included.
• No model weights, logits or FLORES-101 text are included.

Contact

hello@chipsnug.com · Updates: https://chipsnug.com/?utm_source=hf-dataset&utm_medium=tool&utm_campaign=koqloss-v0.2.0

한국어

같은 내용을 두고, 공개 모델을 양자화했을 때 한국어가 영어보다 정보를 얼마나 더 잃는지 잰 결과표다. 원본 저장소는 https://github.com/chipsnug/koqloss이다.

• 같은 내용이면 한국어가 영어보다 1.09~3.45배 더 잃는다(Q8_0 대비, 8조합 중 7조합에서 구간이 1.0 초과).
• 토큰당으로 보면 kanana는 검출할 만한 차이가 없지만(95% 구간 0.841.14), A.X-4.0-Light는 1.171.32배, Qwen3은 1.42~2.04배를 더 잃는다. "한국어 중심 모델은 토큰당 격차가 없다"는 가설은 기각이다.
• 전체 보고서는 koqloss-public.md에 있다.

라이선스는 결과표 CC BY 4.0이다. FLORES-101에서 나온 항목은 CC BY-SA 4.0 조건도 따른다. 모델 가중치·로짓·FLORES-101 원문은 없다.