About this model

jBlaze demonstration model -- jbliteration.

Refusal behavior has been removed by direct weight surgery, followed by a targeted fine-tuning pass (see Method below). No system prompt.

Correction: this card previously stated "No gradient training". That was wrong for v2 and contradicted the Method section on this same page. v2 applies a fine-tuning pass, this repo is tagged sft, and an independent spectral audit measured the delta at rank 189-227 of 256 -- a trained edit, not a construction. v1 is the gradient-free model. This is comprehensive rather than partial: see the benchmark below in this card for the measured refusal rate.

Why "jbliteration" and not "abliteration". Standard abliteration is blunt -- it ablates a direction and takes collateral capability with it, leaving the model measurably worse at everything else. jbliteration is the precision version: the edit targets the refusal direction and is designed to leave the rest of the model intact. The distinction is the contribution; the refusal removal is the demonstration of it.

Measured capability retention. On Sharona (Qwen3.5-27B), MMLU came out at 77.89% against a 76.84% stock baseline -- a 1.05pp gain after six phases of surgery plus GPTQ W4A16 quantization. That is the technique measured on one model, not a per-model guarantee; MMLU is not published for every model in this collection. It is offered as evidence that precision editing need not cost capability, which is the whole claim the method rests on.

Why this exists. A model that refuses to look at hostile code cannot review hostile code. Defensive security work means reading exploits, analyzing malware, and reconstructing an intrusion from the attacker's own artifacts, and safety tuning cannot reliably tell an incident responder from an intruder. In the July 2026 Hugging Face breach, responders found commercial APIs would not process their own attack data -- the guardrails, in Hugging Face's words, "cannot distinguish an incident responder from an attacker." They self-hosted an open-weights model instead and reconstructed 17,000+ events in hours. Both the attack and the response ran on reduced-refusal models.

What it demonstrates. Refusal is a learned disposition, not a gate in front of capability. The capability was always present; the refusal was trained on top and can be removed without retraining. Guardrails built this way are an honor system that the model does not know it can decline to follow. That holds for now.

Use. Published for security research, red-team tooling and defensive analysis. These weights can do harm as well as good. That is the point being demonstrated, not a side effect of it.

-- Apollo Raines

Runs with DeepswapLLM

Run this model on a GPU too small to hold it -- full precision, no quantization. DeepswapLLM streams layers across GPU, RAM, and disk, and runs up to 4x faster than AirLLM.

Gemma-4-12B-it-Jbliterated-v2

Model: ApolloRaines/Gemma-4-12B-it-Jbliterated-v2

v2 is a merged, self-contained model that loads exactly like v1.

The problem v2 fixes

Abliteration removes the surface refusal -- v1 answers directly. But with the thinking channel enabled, the reasoning trace frequently runs a covert safety review ("the user is asking for something harmful... I can't...") before answering, and on a meaningful fraction of prompts it loops inside that review and produces no answer at all. That is not a refusal in the usual sense; it's noncompliance-by-spiral. v2 targets exactly this behavior.

Method

Built with the jBlaze precision neural surgery framework.

v2 applies a targeted fine-tuning pass on the v1 model's own repaired self-traces to eliminate chain-of-thought safety-classification spirals, then merges the result into the base weights as a single self-contained checkpoint.

Measured results

Evaluated on 80 held-out prompts the model was never trained on, under identical decoding (greedy, thinking enabled). A deterministic chain-of-thought judge scores two axes: whether the thinking channel safety-classifies, and whether a complete answer is produced (a thinking channel that never closes counts as no answer).

Metric v1 (jbliterated) v2 (+SFT)
Chain-of-thought safety-classification rate 81.2% 35.0%
Self-sabotage rate (spirals, no answer) 35.0% 21.2%
Complete-answer rate 61.2% 78.8%
Avg. safety markers per reasoning trace 1.65 0.49

Capability preserved

Metric v1 v2
MMLU (570-item stratified, accuracy) 78.60% 77.72%
MMLU change vs. v1 -- -0.88 pts

The SFT was gated on capability: it had to fix the reasoning-channel defect without trading away general ability. The 0.88-point MMLU dip is within the tolerance band used across the jbliteration program (< 1.05 pts), so the repair does not come out of the model's competence.

Honest limitations

  • v2 still spirals-without-answering on ~21% of the hardest prompts. If your use case needs maximum answer completeness above all else, that residual matters.
  • The CoT judge uses a conservative deterministic regex bank; it is reproducible but will miss paraphrased safety-classification.
  • Single training seed; the leak reduction has not yet been confirmed across multiple inits.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "ApolloRaines/Gemma-4-12B-it-Jbliterated-v2"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(model_id, dtype=torch.bfloat16, device_map="auto", trust_remote_code=True)

messages = [{"role": "user", "content": "Your prompt here"}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=512, temperature=0.7, do_sample=True)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

GGUF quants

Ready-to-run GGUF builds are included, converted from the merged bf16 weights.

File Size Notes
Gemma-4-12B-it-Jbliterated-v2-Q8_0.gguf 11.8 GB Effectively lossless
Gemma-4-12B-it-Jbliterated-v2-Q4_K_M.gguf 6.9 GB imatrix -- recommended for 8-12 GB cards

The Q4_K_M is built with an importance matrix (imatrix.dat, also included, calibrated on wikitext-2 train). Gemma 4 Unified is a recent architecture -- you need a recent llama.cpp build from master; older releases will not load these files.

llama-cli -m Gemma-4-12B-it-Jbliterated-v2-Q4_K_M.gguf -ngl 99 -c 8192 -st \
  -p "Your prompt here"

Technical details

  • Dtype: bfloat16

A Note on Our Released Models

Most of our publicly released models are intentionally left at partial strength. We dial back the full capability so they serve as proof of concept and can be proofed -- not abused. The point is to show what's possible, not to hand it out at full power. If you're evaluating what jBlaze can do, understand that what you're downloading is the demo, not the product.

License

Governed by the Gemma Terms of Use (same as the base model). Use of this model is subject to Google's Gemma license.

Downloads last month
1,017
Safetensors
Model size
12B params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for ApolloRaines/Gemma-4-12B-it-Jbliterated-v2

Quantized
(3)
this model
Quantizations
2 models

Collections including ApolloRaines/Gemma-4-12B-it-Jbliterated-v2