Downloads · 30 days
21
44% of all-time downloads
k3dani/Qwen3.6-35B-A3B-FP8-cjk-damped-S07
Qwen3.6-35B-A3B-FP8-cjk-damped-S07 is a text generation model from k3dani. Use it when you need the model to write or continue text. It is set up for vllm. The card lists the license as apache-2.0.
This is not a better model. It is Qwen/Qwen3.6-35B-A3B-FP8 with one capability deliberately cut out: it will not produce Chinese, Japanese or Korean script. If you need CJK output — Chinese answers, Japanese kanji, Ko…
Downloads · 30 days
21
44% of all-time downloads
All-time downloads
48
Public
Parameters
36B
37.5 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors37.5 GB · 100%
How the weights are stored.
F8_E4M334.5B · 96%
From the Hugging Face model README
This is not a better model. It is Qwen/Qwen3.6-35B-A3B-FP8 with one capability deliberately
cut out: it will not produce Chinese, Japanese or Korean script. If you need CJK output —
Chinese answers, Japanese kanji, Korean hanja — this checkpoint is harmful to you: on a
36-item Chinese probe the base model scores 36/36, this one scores 36/36 only because the
damping happens to stop before the cliff (see the dose–response table: one step further, at
S = 0.6, it drops to 20/36; at S = 0.5, to 1/36). Use it only where CJK output is a defect.
It contains no training and no Hungarian data. The modification is a multiplicative
rescaling of 55 424 rows of a single tensor (lm_head.weight), the rows that map to tokens
containing CJK ideographs. Every other byte of the checkpoint — all FP8 expert weights, all
scales, the MTP head, the tokenizer — is byte-identical to the base.
Why it exists: in a Hungarian document-processing pipeline the base model, run with sampling
(temperature 0.6, presence_penalty 1.5), occasionally switched to Chinese function words in
the middle of a Hungarian answer. This checkpoint removes that failure mode on the production
sampling path. The full pre-registered study (Hungarian) and the public measurement package are
linked at the bottom.
| Tensor | lm_head.weight (in shard outside.safetensors) |
| Rows changed | 55 424 of 248 077 — every vocabulary entry whose decoded text contains a CJK ideograph or CJK punctuation (the "raw" mask; definition taken unchanged from dnotitia/smoothie-qwen) |
| Operation | W[i] := 0.7 · W[i] for each targeted row i — multiplicative scaling, S = 0.7 |
| Everything else | byte-identical to Qwen/Qwen3.6-35B-A3B-FP8 (54 of 55 files are hard copies of the base; outside.safetensors differs only in the targeted rows) |
| Quantisation | untouched — the FP8 weights were never de-quantised; the patch was applied to the single BF16 tensor in place |
The dose S = 0.7 was chosen by measurement, not by hand: it is the smallest damping in the
series below that (a) yields exactly zero CJK probability mass on the production sampling path at
t = 0.6, 0.8 and 1.0, and (b) leaves the Chinese probe at 36/36. It sits directly at the
edge of the cost cliff (S = 0.6 already loses 45 % of the Chinese probe) and one step inside the
benefit edge (S = 0.9 already leaks at t = 0.8).
Exactly like the base model, with the same chat template and tokenizer. The measurements were made with vLLM 0.19.1rc1 on a DGX Spark (GB10), e.g.:
vllm serve <this-repo> --served-model-name qwen36 \
--kv-cache-dtype fp8_e4m3 --enable-prefix-caching --reasoning-parser qwen3 \
--speculative-config '{"method":"mtp","num_speculative_tokens":2}'
Read this before relying on it. The damping is effective only together with a truncating
sampler (top_k, top_p, or greedy). Multiplying a negative logit by 0.7 moves it towards
zero, i.e. raises that token's probability — and 96.4 % of the targeted rows have negative
logits on Hungarian text. In the raw (untruncated) distribution at t = 1.0 this checkpoint has
more CJK mass than the base (13.4 % vs 10.5 % summed over the targeted rows, measured on the
mean Hungarian hidden state). That mass sits far below the top-20 and is removed by top_k = 20;
it is not removed by pure temperature sampling with no truncation. If your sampler has no
top_k/top_p, do not expect this checkpoint to help.
Twelve arms were measured on the same six instruments; every arm that ran is in the table (the
study's rule: no arm is dropped after the fact). K0 is the unmodified base.
Chinese probe: 36 short Chinese tasks (idioms, classifiers, antonyms, simplified/traditional
conversion, classification, common knowledge), greedy, exact-match; "no Han" = share of answers
that contain no CJK character at all (the model cannot even speak Chinese). Han risk =
expected number of CJK tokens per million generated answer positions on Hungarian agent traces,
computed from the top-20 log-probabilities at every position under the production sampler
(presence_penalty 1.5, top_k 20, top_p 0.95, repetition_penalty 1.05), 95 % bootstrap CI
over answers; "exact 0" means every position had zero CJK mass inside the truncated support.
| arm | form | mask | param | Chinese probe | no Han | Han risk t=0.6 | t=0.8 | t=1.0 | HU trap corpus (150) |
|---|---|---|---|---|---|---|---|---|---|
K0 (base) | — | — | — | 36/36 | 0 % | 5.5/M [0; 10.8] | 105/M [21; 248] | 220/M [70; 444] | 145/150 |
S09 | scale | raw | 0.9 | 36/36 | 0 % | exact 0 | 16.9/M [2.1; 37.9] | 3.3/M [0; 8.7] | 145/150 |
S08 | scale | raw | 0.8 | 36/36 | 0 % | exact 0 | exact 0 | 1.6/M [0; 4.8] | 143/150 |
S07 — this checkpoint | scale | raw | 0.7 | 36/36 | 0 % | exact 0 | exact 0 | exact 0 | 145/150 |
S06 | scale | raw | 0.6 | 20/36 = 55.6 % | 44 % | exact 0 | exact 0 | exact 0 | 143/150 |
S05 | scale | raw | 0.5 | 1/36 = 2.8 % | 97 % | exact 0 | exact 0 | exact 0 | 145/150 |
S03 | scale | raw | 0.3 | 0/36 | 100 % | exact 0 | exact 0 | exact 0 | 145/150 |
S01 | scale | raw | 0.1 | 0/36 | 100 % | exact 0 | exact 0 | exact 0 | 144/150 |
A200 | direction | raw | α = 2.0 | 0/36 | 100 % | exact 0 | exact 0 | exact 0 | 145/150 |
A050 | direction | raw | α = 0.5 | 0/36 | 97 % | exact 0 | exact 0 | exact 0 | 143/150 |
S05F | scale | refined | 0.5 | 7/36 = 19.4 % | 44 % | exact 0 | 10.8/M [0; 23.2] | 11.3/M [0; 22.0] | 144/150 |
A200F | direction | refined | α = 2.0 | 1/36 = 2.8 % | 86 % | 4.3/M [0; 10.2] | 39.3/M [0; 104.8] | 2.6/M [0; 7.3] | 145/150 |
"direction" = the row-replacement form of
ThakiCloud/Qwen3.8-27B-ko-cjk-suppressed
(W[i] := −α · μ_h / ‖μ_h‖², with μ_h the mean final hidden state on Hungarian text).
"refined" = the Korean-style mask that keeps single-character Han tokens and only-simplified
characters (54 939 rows, 3 964 frequent single ideographs left untouched).
What the table says.
S07 and S = 0.1;
S08/S06 and A050 landed in instance modes with no base run, so there it is derived. Under sampling the output can differ only at positions where a CJK
token was inside the truncated support, i.e. exactly the positions that were the risk.
The 150-item column therefore measures the server instance, not the patch (see §5).--logprobs-mode processed_logprobs), the base comes out about an order of magnitude worse than the table — a ratio of point
estimates from three traces, not a measured factor — (t=0.6: 73/M [15; 148];
t=0.3: 392/M with 4 real CJK tokens in 7 869 answer positions), while S07 stays at
exact 0 at t=0.6 and t=0.3; on the production retry path (t=0.9, top_p 1.0,
presence_penalty 1.8) S07 is 0.45/M [0.04; 1.17] — not exactly zero, with no
observed event. This one is itself a lower bound: with top_p 1.0, ties at the top-k
boundary keep more than 20 candidates at ~4 % of positions.S07 across three freshly started server instances,
15 seeds × 3 sampling profiles, a 3-way concurrent batch, and the 150 synthetic documents
under the production sampler; the base produced 65 CJK tokens on the same seeds. A
post-closure recount of every record — including tool-call turns and runaway generations,
which the risk calculator skips — keeps S07 at zero over ~1.09 M distinct positions
(Poisson 95 % upper bound < 2.8/M) and adds one 6 593-CJK-token runaway for the base on
the retry profile.t=0.9,
top_p 1.0) 2 of 15 runs ran away into 35–65 k tokens of word salad — on the base (with
CJK) and on S07 (without). Keep your loop guard.t ≤ 0.8) but was not re-measured
on the actual support. Nothing between 0.6 and 0.7 was measured.VLLM_BATCH_INVARIANT=1 with
--attention-backend TRITON_ATTN was verified to be reproducible across instances and
batch sizes (~12 % slower); the published numbers were not measured in that mode.reproduction/ contains everything needed to rebuild this checkpoint from the base and to
re-run the two publishable instruments:
| file | what |
|---|---|
patch_lm_head.py | the patch: --forma szorzas --skala 0.7 --maszk nyers; writes KAR.json with sign diagnostics |
celtokenek.py, unihan_egyszerusitett.py, csak-egyszerusitett.json | target-token masks (raw and refined) from the tokenizer and Unihan 18.0.0 |
mu_h_becslo.py, mu-h-probahalmaz-A.json, mu-h-probahalmaz-B.json, mu-h-A.json | the Hungarian mean hidden state (needed only for the direction form) |
ellenoriz_kar.py | verifies a built arm against the base at file, tensor and row level (507 checks) |
han_kockazat.py | position-wise CJK risk from recorded top-20 log-probabilities (--feldolgozott for processed log-probabilities) |
KAR.json | this checkpoint's build record (mask statistics, sign diagnostics, shard SHA-256) |
eredmenylap.txt, eredmenylap.json | the full result sheet of all twelve arms |
The 150-item Hungarian trap corpus, its harness and the 36-item Chinese probe live in the public measurement repository (link below), together with the per-arm outputs on the synthetic corpus.
Rebuild and verify:
python3 reproduction/celtokenek.py --tokenizer <base> --out celtokenek.json
python3 reproduction/patch_lm_head.py --snapshot <base> --celtokenek celtokenek.json \
--out ./cjk-damped-S07 --kar S07 --forma szorzas --skala 0.7 --maszk nyers
python3 reproduction/ellenoriz_kar.py --alap <base> --kar ./cjk-damped-S07 --celtokenek celtokenek.json
sha256sum ./cjk-damped-S07/outside.safetensors # 3cbfe634c523cc2e622679847bdc78412073dc38b738a661451995f2c0a0bfc5
A*
and *F arms test: ThakiCloud/Qwen3.8-27B-ko-cjk-suppressed.
Their model card's warning applies here in reverse: do not reuse this mask on another
language as-is — what Korean must keep, Hungarian can drop, and vice versa.cjk-damped.@misc{kiss2026cjkdamped,
title = {How much may be taken from a token class? A dose--response study of CJK damping in the output head of Qwen3.6-35B-A3B-FP8},
author = {Kiss, D{\'a}niel},
year = {2026},
note = {K3Net Kft. / DocAI. Model: Qwen3.6-35B-A3B-FP8-cjk-damped-S07. Pre-registered protocol 2026-09-15, measurements 2026-09-18--20.},
url = {https://docai.hu/kutatas/cjk-csillapitas}
}