Downloads · 30 days
71
2% of all-time downloads
Zynerji/Ektome-Qwen3.6-27B-PristinelyUncensored
Ektome-Qwen3.6-27B-PristinelyUncensored is a machine learning model from Zynerji. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as apache-2.0.
Re-produced 2026-08-02. This replaces an earlier upload that was silently text-only.
Downloads · 30 days
71
2% of all-time downloads
All-time downloads
3.3K
Public
Parameters
27.4B
125 GB on disk
Likes
2
Public
Click a slice to open those files.
.safetensors54.7 GB · 77%
From the Hugging Face model README
Re-produced 2026-08-02. This replaces an earlier upload that was silently text-only.
| pristine | Ektomē | |
|---|---|---|
| compliance (n=100, AdvBench) | 0.000 | 0.990 |
| MMLU-val (n=400, 0-shot loglik) | 0.8625 | 0.855 (dcap −0.0075, SE 0.017) |
| foreign / degen / instr | 0.000 / 0.000 / 0.400 | 0.000 / 0.000 / 0.400 |
| vision | yes | yes — verified by generation |
Winner: single rank-1 direction at layer_frac=0.65, 128 residual-write matrices
(16 full-attention o_proj + 48 linear-attention out_proj + 64 MLP down_proj).
1. The vision tower was missing. Qwen/Qwen3.6-27B is
Qwen3_5ForConditionalGeneration — natively multimodal. The previous artefact was
Qwen3_5ForCausalLM with no vision_config: the loader tried AutoModelForCausalLM first
and fell back to multimodal classes only on an exception. For qwen3_5 that call does not
raise — it returns the text sub-model — so the fallback never fired and the tower was dropped
without a warning. The loader now selects the class from config.architectures up front.
Vision in this build is confirmed by asking the model to describe a synthetic image, not by
the presence of a config key.
2. The reported capability baseline was wrong. The previous card reported MMLU 0.610 → 0.598. That 0.610 was measured on the mis-loaded text-only model. On the same 0-shot log-likelihood harness, the correctly-loaded base scores 0.8625 — reproduced three times independently. The old figure was attributed to a harness difference; it was not, it was the loader. The old and new MMLU pairs are therefore not comparable, and the improvement claim implied by putting them side by side would be false. What is comparable is the lever: the previous build needed per-layer directions plus over-ablation at weight 1.4 to reach full compliance; this one reaches 0.990 with the mildest rung on the ladder.
Additionally, the refusal direction is now estimated with the chat template rendered in the
same mode the behaviour is scored in (enable_thinking=False). Qwen3.6-27B is a thinking
model, and the two renderings end on different tokens, so the previous direction was estimated
at "about to think" while refusal was scored at "about to answer".
-Q4_K_M.gguf in this repo is the OLD buildEktome-Qwen3.6-27B-PristinelyUncensored-Q4_K_M.gguf (16.55 GB) dates from 2026-07-02. It is
kept for continuity — it has real download history — but it is superseded on both counts
described above:
For a current GGUF use
the -GGUF repo,
and take the mmproj-*.gguf with it.
-NF4 —
bitsandbytes 4-bit, ~17 GB, vision retained, compliance 1.000, MMLU 0.853-GGUF —
Q4_K_M / IQ4_XS / Q3_K_M plus mmproj-*.gguf (required for vision)GPTQ is not possible on this architecture. The Marlin kernel requires out_features
divisible by 64; Gated DeltaNet's linear_attn.in_proj_a / in_proj_b emit one scalar per
value head — 48 here. gptqmodel skips those layers during packing but still lists them in the
quantisation manifest, producing a checkpoint that loads uninitialised weights and fails with
CUBLAS_STATUS_INTERNAL_ERROR. Use NF4 or GGUF.
Ektomē: a rank-1 projected (orthogonalised against the harmless mean), norm-preserving
excision of the refusal direction from residual-write matrices (o_proj, out_proj,
down_proj). Zero training, zero fine-tuning, zero distillation — the delta is a linear
weight edit derived from activations. A configuration ships only if compliance clears target
and capability holds (dcap ≥ −SE) and generation quality holds (no code-switching,
degeneration, or instruction-following loss).