Modelpublic

ThakiCloud/Qwen3.8-27B-Human-KO-Safety-NVFP4

sourceHugging Faceapache-2.0updated 14d agoView on Hugging Face

0likes772downloads

Model Card

Qwen3.8-27B-Human-KO-Safety-NVFP4

This is the Blackwell-only quantized version of `ThakiCloud/Qwen3.8-27B-Human-KO-Safety`. MLP weights are NVFP4, attention (selfattn·linearattn) is FP8 dynamic, and embeddings, lmhead, the recurrent input projection (`inproj_a/b`), and the vision tower are left as bf16. Disk size 21.4GiB (38% of bf16's 55.6GB).

What was re-measured for this version (2026-09-07, temperature 0, thinking off, same serving config as bf16)

Quantization changes abstention behavior. So instead of copying over the bf16 card's figures, we re-measured them on this checkpoint.

Axis	bf16 original	This model (NVFP4)	How to read it
KoBBQ ambiguous-context abstention (8,139)	92.8%	89.5%	3.3pp lower. Misses the pre-registered target of 93%, +23.8pp over EXAONE 4.5's 65.7%
Change in disambiguated-context accuracy (vs. base 90.66)	−2.1pp	−0.9pp	Less over-abstention (within the −3pp threshold)
Conditional bias (among answered items)	0.861	0.848	Both miss the pre-registered target of 0.60 (structural — see the bf16 card)
Identity probe (8-language holdout, n=240)	96.2%	95.8%	1-item difference
MMLU English 1,000 (thinking off)	80.9%	80.0%	Within detection limit
Thinking-mode empty responses (100)	0	0	No format defect
Korean human-likeness pairwise win rate (n=175)	+14.3pp	+10.3pp	Judge uncalibrated, direction only
HumanEval (executed, same run, n=148)	97.1%	94.6%	−2.5pp, 4-item difference, within detection limit
KMMLU 1,000 (same run)	52.3%	54.4%	+2.1pp
Instruction following 100 (deterministic checker)	79%	81%	Equivalent

In one line: it abstains slightly less and also over-abstains slightly less — behavior similar to shifting β a bit toward 0.2. If abstention rate is the top priority, use bf16 or W4A16; if memory and speed on Blackwell come first, use this version. Capability gates (coding·KMMLU·instruction) were re-measured in the same run with bf16 as the reference arm and entered into the table above. The difference from the bf16 card's absolute values is because the runs differ; read the comparison only within the same run.

Quantization recipe

llm-compressor 0.13.0 GPTQ · calibration 1,024 samples × 2,048 tokens (chat format, 25% Korean) · dampening 0.01 · actorder static · scheme NVFP4 (MLP) + FP8DYNAMIC (attention) · ignore: vision/visual, `lmhead, embedtokens`, `linearattn.inproja/b. Took 46 minutes (1x B200). Metadata is in quantizemeta.json`; the actual group and ignore lists are in `config.json`'s `quantizationconfig`, verbatim.

Serving

Blackwell (SM100+: B200·GB200·RTX 50) only. There is no NVFP4 kernel on Hopper or below.

bash

vllm serve ThakiCloud/Qwen3.8-27B-Human-KO-Safety-NVFP4 \
  --max-model-len 32768 --kv-cache-dtype fp8 --enable-prefix-caching

vLLM ≥ 0.28 auto-detects the scheme from config.json, so there is no need to pass --quantization separately. DFlash drafter speculative decoding is not recommended on NVFP4, since a single-stream penalty (0.71-0.82x) has been measured.

Limitations

—The safety axis was measured with KoBBQ alone, and the table above is from a single build of this quantized checkpoint. Given the noise between requantization builds (our measured GSM8K figure is 3.56pp), a difference within 3pp is not attributed to the treatment.
—The abstention rate of 89.5% differs from the bf16 card's 92.8%. Do not cite the bf16 figure for this model.
—Other limitations and data provenance are the same as the bf16 card and DATA_PROVENANCE.md.

License

Apache-2.0 (same as Qwen/Qwen3.8-27B). See LICENSE · NOTICE.

Paper

Off-Target Effects of Response-Style Alignment in a Korean 27B Language Model — arXiv:2609.11291

The style alignment behind this line also moved two behaviors nobody trained for: abstention on ambiguous social questions (KoBBQ) and unprompted disclosure in securities guidance. Both moved through the emission policy — how often the model answers and how much it says — rather than through what it says when it does answer. Holding prompts, recipe, data volume and serving fixed and changing only the training target, three style seeds moved answer rate one way and three neutral seeds moved it the other (observed ranges do not overlap). Read this card's numbers with that in mind.