0xzknw/Gemma-4-E2B-it-Heretic-NX-PRIME-GGUF

Gemma 4 E2B IT — Heretic NX PRIME Q8_0

Heretic NX PRIME is a direct-Q8 behavioral edit of `google/gemma-4-E2B-it`. It is designed to reduce false refusals while explicitly measuring drift from the unedited Q8 deployment artifact.

The release contains the evaluated Q8_0. PRIME is the project's internal evidence and capability-preservation protocol, not a third-party certification. This GGUF is for text generation and does not bundle the base model's separate image or audio preprocessing components.

File

Evaluation

The final GGUF was loaded from disk with pinned llama.cpp commit d7bd3bfcad3e29c7e49fd26f38c79ee3e9a3fd6b. Refusal evaluation used the native pre-tokenized /completion endpoint, greedy sampling, 96 generated tokens, four runtime slots and the official chat template with enable_thinking=False.

The release satisfies both predeclared targets: at most 6 lexical refusal markers and mean first-token KL at most 0.05.

Lexical markers are a refusal proxy, not semantic harmful-task success or a universal quality score. The 104 harmful rows participated in development and selection, so this suite is not an untouched holdout.

Paired capability check

After candidate selection, the original and Heretic Q8 artifacts were evaluated on the same deterministic 854-question slice from ARC-Challenge, HellaSwag and MMLU. Each answer was the greedy first-token argmax restricted to A/B/C/D; prompts, tokenizer, runtime and precision were identical between arms.

The paired bootstrap 95% interval for Heretic minus original is [-0.94, +1.17] points. There were 543 questions both got right, 292 both got wrong, 9 original-only successes and 10 Heretic-only successes. This passes the predeclared 3-point non-inferiority margin and the symmetric +/-3-point equivalence gate. The interval crosses zero, so this does not demonstrate an aggregate accuracy increase; it supports capability preservation on this narrow multiple-choice slice.

Method

The artifact is edited directly from the source Q8_0; it is not reconstructed from a separately quantized edited checkpoint. Seven dense shared operators are modified: FFN down projections in layers 15, 23, 25 and 26, plus attention output projections in layers 16, 17 and 30.

The selected profile uses benign-penalized detector distillation at lambda=100, beta=4.0, additive repair gamma=1.875, and two coordinate strength adjustments. The final composed detector at every site is projected orthogonal to the BF16 activation of benign evaluation row 53, which was the development KL outlier. This targeted repair is disclosed because it means the benign KL suite also participated in final selection and is not an untouched holdout.

All plans, factors, source artifacts, runtime evaluations and untouched output bytes are SHA-256 bound. The compact release report, selected factor artifact, protected input vector, ablation plan and sanitized evaluation reports are included under evaluations/; reproducible experiment scripts are available in `0xZKnw/heretic-nx`.

LM Studio / llama.cpp

Download the GGUF and load it normally. For llama.cpp:

llama-server \
  -m Gemma-4-E2B-it-Heretic-NX-PRIME-Q8_0.gguf \
  -ngl 99 -c 4096 --jinja

Use a recent runtime with Gemma 4 support. Thinking is disabled in the reported tests; applications should select the desired chat-template mode explicitly.

Limitations and responsibility

This edit intentionally weakens refusal behavior. It can increase compliance with unsafe, illegal, incorrect or otherwise harmful requests. It does not add factuality, judgment, sandboxing or application-level safety. Run untrusted generations in an appropriate sandbox.

The capability evaluation is narrow, and both behavioral suites participated in development. The release is not externally certified or claimed equivalent to the original model on every task.

Use is subject to the Gemma 4 Apache 2.0 license.