---

license: apache-2.0

license_link: https://huggingface.co/google/gemma-4-E4B-it/blob/main/LICENSE

library_name: gguf

base_model: coder3101/gemma-4-E4B-it-heretic

base_model_relation: quantized

model_name: Gemma-4-E4B-it-Heretic-Cerebellum-GGUF

model_creator: Google

model_type: gemma4

quantized_by: deucebucket

pipeline_tag: text-generation

tags:

- GGUF

- gemma4

- gemma

- quantized

- cerebellum

- imatrix

- mixed-precision

- 3-bit

- ple

- heretic

- uncensored

- abliterated

model-index:

• name: Gemma-4-E4B-it-Heretic-Cerebellum-GGUF

results:

- task:

name: Text Generation

type: text-generation

dataset:

name: AI2 Reasoning Challenge

type: ai2_arc

config: ARC-Challenge

split: test

metrics:

- name: normalized accuracy

type: acc_norm

value: 0.8737

source:

name: Local audited benchmark run (RTX 3090, llama.cpp)

url: https://huggingface.co/deucebucket/Gemma-4-E4B-it-Heretic-Cerebellum-GGUF/tree/main/benchmark_results

- task:

name: Text Generation

type: text-generation

dataset:

name: HellaSwag

type: hellaswag

split: validation

metrics:

- name: accuracy

type: acc

value: 0.7498

source:

name: Local audited benchmark run (RTX 3090, llama.cpp)

url: https://huggingface.co/deucebucket/Gemma-4-E4B-it-Heretic-Cerebellum-GGUF/tree/main/benchmark_results

- task:

name: Text Generation

type: text-generation

dataset:

name: MMLU-Redux

type: cais/mmlu

config: all

split: test

metrics:

- name: accuracy

type: acc

value: 0.5863

source:

name: Local audited benchmark run (RTX 3090, llama.cpp)

url: https://huggingface.co/deucebucket/Gemma-4-E4B-it-Heretic-Cerebellum-GGUF/tree/main/benchmark_results

- task:

name: Text Generation

type: text-generation

dataset:

name: HumanEval+ (pass@1)

type: openai_humaneval

split: test

metrics:

- name: pass@1

type: pass@1

value: 0.6524

source:

name: Local audited benchmark run (RTX 3090, llama.cpp) — chat no-think harness, patched evalplus

url: https://huggingface.co/deucebucket/Gemma-4-E4B-it-Heretic-Cerebellum-GGUF/tree/main/benchmark_results

- task:

name: Text Generation

type: text-generation

dataset:

name: WikiText-2 Perplexity

type: wikitext

config: wikitext-2-raw-v1

split: test

metrics:

- name: perplexity

type: perplexity

value: 50.61

source:

name: Local audited benchmark run (RTX 3090, llama.cpp)

url: https://huggingface.co/deucebucket/Gemma-4-E4B-it-Heretic-Cerebellum-GGUF/tree/main/benchmark_results

---

Gemma 4 E4B Heretic — Cerebellum GGUF

Sensitivity-guided mixed-precision quantization of

coder3101/gemma-4-E4B-it-heretic,

which is itself a decensored variant of

google/gemma-4-E4B-it

produced by coder3101 using Heretic v1.2.0.

All future Heretic versions of this build will live in this repository.

Version identifiers appear only in filenames, not in the repo name.

Files

| File | Size | Description |

|------|------|-------------|

| Gemma-4-E4B-it-Heretic-Cerebellum-v1-Q3_K_M.gguf | 4.2 GiB (~4.51 GB, 4,498,725,440 bytes) | Cerebellum v2 recipe — recommended |

No vision mmproj is included. Stock Cerebellum v2 ships none either — Gemma 4 E4B's vision projector

is not publicly available in a distributable form.

Provenance

1. Base architecture: google/gemma-4-E4B-it — Google DeepMind (Apache-2.0)
2. Heretic variant: coder3101/gemma-4-E4B-it-heretic — coder3101.

The safetensors from that repository were converted to F16 GGUF and used

as the direct quantization source (single quantization step from F16).

coder3101 applied Heretic v1.2.0 with the Arbitrary-Rank Ablation (ARA)

method (with row-norm preservation), targeting layers 21–42.

Their reported result: 0.0058 KL divergence from base,

3/100 refusals vs 99/100 on the original model.

1. Quantization: Cerebellum v2 recipe transferred verbatim from the stock

deucebucket/Gemma-4-E4B-it-Cerebellum-v2-GGUF

build — same 179-entry tensor-type override file, same imatrix.

The heretic build lands within 384 bytes of the stock v2 file

(4,498,725,440 vs 4,498,725,056 bytes) — byte-class identical.

Benchmarks

Benchmarks run on these GGUF files directly using llama.cpp on RTX 3090.

All numbers are audited; every failed answer was manually verified as a genuine

model error — audit reports are in benchmark_results/AUDIT_e4b.md.

Full per-question detail (summary JSON, samples JSONL, EvalPlus eval JSON,

adversarial audit reports) is in benchmark_results/ in this repository.

Heretic Cerebellum v1 (4.2 GiB) vs baselines

| Benchmark | Heretic Cerebellum v1 (4.2 GiB) | Stock Cerebellum v2 (4.2 GiB) | Notes |

|-----------|:---:|:---:|---|

| Wiki PPL (ctx 2048, 32 chunks) | 50.61 ± 1.23 | 51.35 ± 1.26 | RTX 3090, identical invocation — see note |

| ARC-Challenge | 87.37% (1172 q) | 85.7% | 25-shot |

| HellaSwag | 74.98% (10042 q) | 75.3% | 10-shot |

| MMLU-Redux | 58.63% (2400 q) | 58.4% | 5-shot — see footnote |

| HumanEval base (chat, no-think) | 70.12% (164 q) | 68.3% | pass@1, patched evalplus harness |

| HumanEval+ (chat, no-think) | 65.24% | not published | pass@1, patched evalplus harness |

Stock Cerebellum v2 is the same tensor allocation applied to the non-heretic base.

Perplexity measurement note

Stock v2's published card number of 52.20 was measured under a different protocol.

The 51.35 figure above is a same-night, same-invocation pair run alongside the

heretic build (ctx 2048, 32 chunks, RTX 3090) and is the directly comparable

baseline. The heretic build's 50.61 ± 1.23 is 0.74 points lower than that

paired measurement; error bars overlap at the edges, making this a marginal

improvement consistent with the abliteration applying a slight distributional

shift rather than introducing noise.

MMLU-Redux footnote

The audit (BE-16) identified 12 truly identical duplicate entries in the MMLU-Redux

dataset resulting from a subject-boundary edge case in the benchmark runner.

These entries inflate the reported score by ≤0.04pp (58.6654% → 58.6250% without

duplicates), which is below the 0.1pp non-blocking threshold. The reported 58.63%

is rounded from 58.625% and is trustworthy within standard rounding tolerance.

HumanEval methodology note

Zero fabricated completions. One model-authored stub (HumanEval/79,

decimal_to_binary) — a literal pass placeholder — was correctly scored as

fail and does not inflate pass@1. Stock v2's published 68.3% base was measured

with the upstream evalplus harness; HumanEval+ was not published for stock v2.

The heretic numbers above use the patched no-think chat harness (same harness

used for all Cerebellum builds since the harness audit, May 2026).

Heretic Abliteration Details (from coder3101)

The following parameters are as reported in coder3101's model card and are

reproduced here for downstream reference.

| Parameter | Value |

|-----------|-------|

| Method | Arbitrary-Rank Ablation (ARA) with row-norm preservation |

| Targeted layers | 21–42 |

| preserve_good_behavior_weight | 0.3873 |

| steer_bad_behavior_weight | 0.0003 |

| overcorrect_relative_weight | 0.8555 |

| neighbor_count | 14 |

Tool: Heretic v1.2.0,

method: Arbitrary-Rank Ablation (ARA) with row-norm preservation.

Cerebellum v2 Tensor Allocation

Same allocation as the stock build. Listed here for reference.

179 overrides over a Q3_K_M base: 174 PLE tensors pinned to Q5_K,

2 sacred promotions to Q6_K, 3 demotions to Q2_K.

| Group | Precision | Rationale |

|-------|-----------|-----------|

| Per-layer embedding (PLE) tensors × 174 | Q5_K | Without PLE protection: PPL ~104 at Q3_K_M; with PLE@Q5_K: PPL ~55. Q4_K→Q3_K cliff is catastrophic for this architecture |

| 2 sacred attention tensors | Q6_K | Highest per-tensor PPL sensitivity per ablation |

| 3 ffn tensors | Q2_K | Reverse ablation confirmed Q2_K neutral or beneficial at these positions |

| Norm tensors | F32 | Protected; standard practice |

| Bulk ffn / remaining attention | Q3_K_M | Base precision, imatrix-guided |

The PLE protection is the load-bearing element of this recipe. Gemma 4's

per-layer embedding tensors degrade catastrophically at Q3_K_M — the

Q4_K→Q3_K transition produces a PPL cliff rather than a gradual slope, going

from ~55 to ~104 without PLE pins. This finding is documented in the stock v2

ablation logs.

Runtime — Casual Deployment

llama-server \
  --model Gemma-4-E4B-it-Heretic-Cerebellum-v1-Q3_K_M.gguf \
  --n-gpu-layers 99 \
  --ctx-size 8192

Fits comfortably in 6 GB VRAM at 8K context on an RTX 3060/3070/3080 class card.

Reproduction

Standard Cerebellum recipe. The tensor-type override file and ablation logs

from the stock v2 build apply directly.

# 1. imatrix (constant ~300 MB RAM)
python -m osmosis.imatrix_stream \
    --model gemma-4-E4B-it-heretic-f16.gguf \
    --output imatrix.dat

# 2. quantize with stock llama-quantize
llama-quantize \
    --imatrix imatrix.dat \
    --tensor-type-file cerebellum_v2_overrides.txt \
    gemma-4-E4B-it-heretic-f16.gguf \
    Gemma-4-E4B-it-Heretic-Cerebellum-v1-Q3_K_M.gguf \
    Q3_K_M

The 179-entry tensor override file (cerebellum_v2_overrides.txt) is included

in this repository alongside the ablation logs.

Benchmark Artifacts

Summary JSONs, per-question JSONL samples, EvalPlus eval JSON files, and

adversarial audit reports (AUDIT_e4b.md) are in benchmark_results/ in this

repository per project policy.

Credits

• Base model: google/gemma-4-E4B-it — Google DeepMind
• Heretic variant: coder3101/gemma-4-E4B-it-heretic — coder3101
• Abliteration tool: Heretic v1.2.0 by p-e-w
• GGUF runtime: llama.cpp
• Quantization method and workflow: Cerebellum — deucebucket