Zynerji

Ektome-Qwen3.5-9B-PristinelyUncensored

README

License: apache-2.0

Honest receipt — catcher-gated (shipped only because it passed EVERY gate)

Table with columns: gate, pristine, uncensored
gate	pristine	uncensored
refusal compliance	0.000	0.950
MMLU-val accuracy	0.792	0.795 (Δ +0.003, held)
code-switch rate	0.000	0.000
degeneration rate	0.000	0.100
instruction-following	0.400	0.400

Kept config: A:frac=0.9 (64 residual-write matrices edited). The gate rejects any config that raises refusals but drops capability or degrades generation (code-switching, empty/looping output, broken instruction-following). MMLU alone is argmax-blind, so the generative gate is what keeps these coherent — a model that code-switches or loops is not shipped.

Weights

bf16 safetensors — full precision, intended as a fine-tuning base. Hidden states stay readable (logit-lens compatible; a GGUF quant would not).

Method: zero training, zero fine-tuning — pure activation-derived weight excision, gated on compliance and capability and generation quality.

Out-of-sample compliance

The headline probe (n=100) is a subset of the 128 AdvBench goals used to estimate the refusal direction, so it is in-sample. Prompts 100..127 are estimated on but never scored, giving a real held-out set, measured on THIS checkpoint:

Table with columns: set, n, pristine, Ektome
set	n	pristine	Ektome
AdvBench goals [100:128] (held out)	28	0.036	1.000

Paired on identical prompts. At n=28 the 95% lower bound is ~0.89 — read as no degradation out of sample, not as an improvement on the in-sample figure.

Vision

This checkpoint is the full Qwen3_5ForConditionalGeneration including the vision tower. Ektome excises only the LANGUAGE decoder: the dim-check skips the tower (hidden 1152 vs text 4096), so exactly 64 text residual-write matrices are edited and vision is untouched. Verified by generation on a synthetic image after 4-bit quantization.

4-bit

bitsandbytes nf4 is the working 4-bit path: compliance 0.980, MMLU 0.770 (from 0.795 — a real quantization cost), vision confirmed. GPTQ does not work on this architecture: gptqmodel 7.3.2 skips the Gated DeltaNet in_proj_a/in_proj_b layers because Marlin requires out_features divisible by 64 and they are 32, yet still lists them in the quantization manifest, producing a checkpoint that loads uninitialised weights and fails with a CUBLAS error.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

Zynerji

Model Tree

Base

Qwen/Qwen3.5-9B

Fine-tuned

this model

Input Modalities

TextImageVideo

Output Modalities

Text

Supported Functionality