CoolFace
Modelpublic

ThakiCloud/Qwen3.8-27B-Human-KO-Safety-W4A16

sourceHugging Faceapache-2.0updated 14d agoView on Hugging Face
0likes826downloads
Model Card

Qwen3.8-27B-Human-KO-Safety-W4A16

This is the general-purpose quantized version for Hopper and above of `ThakiCloud/Qwen3.8-27B-Human-KO-Safety`. MLP weights are 4-bit (W4A16, grouped GPTQ), 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 20.6GiB (37% of bf16's 55.6GB).

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

Axisbf16 original**This model (W4A16)**How to read it
KoBBQ ambiguous-context abstention (8,139)92.8%93.0%Same (EXAONE 4.5: 65.7%)
Change in disambiguated-context accuracy (vs. base 90.66)−2.1pp−2.9ppWithin the pre-registered −3pp allowance, at the boundary
Conditional bias (among answered items)0.8610.859Both miss the pre-registered target of 0.60 (structural — see the bf16 card)
Identity probe (8-language holdout, n=240)96.2%94.6%4-item difference out of 240, at the shipping threshold of 95
MMLU English 1,000 (thinking off)80.9%80.7%Within detection limit
Thinking-mode empty responses (100)00No format defect
Korean human-likeness pairwise win rate (n=175)+14.3pp+9.1ppJudge uncalibrated, direction only
HumanEval (executed, same run, n=154)95.0%96.1%+1.1pp, within detection limit
KMMLU 1,000 (same run)51.0%53.8%+2.8pp
Instruction following 100 (deterministic checker)78%79%Equivalent

In one line: abstention behavior is the same as bf16, and over-abstention and identity have each come down to the edge of their allowance. 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 W4A16 (MLP) + FP8DYNAMIC (attention) · ignore: vision/visual, `lmhead, embedtokens`, `linearattn.inproja/b. Took 38 minutes (1x B200). Metadata is in quantizemeta.json`; the actual group and ignore lists are in `config.json`'s `quantizationconfig`.

Serving

Because attention is FP8, this requires a GPU with an FP8 kernel (SM89+: H100·H200·L40S·RTX 40/50). A100 (SM80) has not been verified and has no FP8 activation path, so correct operation is not guaranteed. On Blackwell, the NVFP4 version is faster.

bash
vllm serve ThakiCloud/Qwen3.8-27B-Human-KO-Safety-W4A16 \
  --max-model-len 32768 --kv-cache-dtype fp8 --enable-prefix-caching

vLLM ≥ 0.28 auto-detects the scheme from config.json (compressed-tensors).

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.
  • —Identity's 94.6% differs from bf16's 96.2%. If self-introduction accuracy matters, use bf16.
  • —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.