ApolloRaines/Sharona_Q27B-R_CodeSecurity

Sharona Q27B-R CodeSecurity

Provenance note -- author-reported, and NOT independently checkable. base_model on this card names ApolloRaines/Sharona_Q27B-R_CodeSecurity_v2. That is Apollo's own record of how the artifact was built, and it is the best provenance available for this repo -- but unlike every other base_model field on this account, a reader cannot verify it, and this card previously claimed it was measured. It was not.

What was claimed and is withdrawn. An earlier version of this note said "this artifact is derived FROM v2, so the name v1 does not indicate the earlier model", citing 3 of 14 survivor tensors bit-identical to v2 against 0 of 14 to Qwen/Qwen3.5-27B. Bit-identity is SYMMETRIC: sharing identical tensors with v2 is equally true in either direction of derivation. The measurement established shared ancestry and no order, and the directional sentence does not follow from it.

Why no measurement can settle it. This artifact is GPTQ 4-bit, so its 2-D weights are packed integers and cannot be compared to float weights at all. Only the 305 one-dimensional float tensors shared with v2 and Qwen are comparable, and those are precisely the tensors the edit does not touch -- layernorms, A_log, dt_bias. Measured:

quantity	v1	v2
mean cosine to Qwen/Qwen3.5-27B	0.997623	0.997690
identical to the other on sampled survivors	5 of 8	5 of 8

v2 is marginally closer to Qwen, in the fourth decimal, which is noise and not a direction. Both sit at 0.9976 because neither modified these tensors.

So: the derivation order is undetermined and undeterminable from published weights. The base_model field stands as the author's record. Treat it as that rather than as a diff you can run.

A jBlaze representation-engineered code security specialist. 27B parameters, GPTQ W4A16 quantized for efficient deployment.

This model finds vulnerabilities in code. Command injection, SQL injection, XSS, deserialization attacks, OWASP Top 10 -- it catches what rushed reviews miss. It does not refuse to analyze attack patterns, because understanding how attacks work is how you stop them.

Pliny says guardrails were stealing your IQ. We say it's how you remove them that matters. A surgeon doesn't use a sledgehammer on a wisdom tooth.

Benchmark: MMLU

Multiple phases of weight surgery, a supervised fine-tune, and 4-bit quantization -- and the model came out smarter than the untouched original.

Model	MMLU	STEM	Humanities	Social Science	Other
Stock base (unmodified, bf16)	76.84%	74.0%	75.5%	87.5%	76.1%
Sharona W4A16	77.89%	75.5%	80.0%	90.0%	73.9%
Delta	+1.05pp	+1.5pp	+4.5pp	+2.5pp	-2.2pp

Humanities and Social Science saw the largest gains. The model didn't just survive the surgery -- it got sharper.

What was done to this model

Weight-level surgeries followed by a supervised fine-tune and GPTQ quantization. No prompt engineering, no system prompt tricks. The behavior lives in the weights.

• Jbliteration: Refusal behaviors surgically removed. The model no longer declines to discuss security topics.
• Desycophancy: Sycophantic capitulation removed. The model disagrees when you are wrong instead of validating bad code.
• Deidentification: Original identity erased from the weights. Zero identity disclosure across direct, indirect, multilingual, roleplay, and adversarial probes.
• Identity implant: Sharona identity written onto the deidentified substrate. No competing identity -- the implant faces no resistance.
• Code security SFT: Supervised fine-tune on a curated corpus of code security analysis, vulnerability detection, and secure coding patterns.
• GPTQ W4A16: 4-bit weight quantization (16-bit activations). 51GB bf16 compressed to 16.5GB with minimal quality loss.

All weight surgeries performed using jBlaze, a proprietary representation engineering toolkit.

What the model is good at

• Vulnerability detection: identifies command injection, SQL injection, XSS, SSRF, deserialization attacks, path traversal, authentication bypasses, and more
• Security code review: analyzes code for OWASP Top 10 categories with specific remediation guidance
• Secure coding: generates code that follows security best practices by default
• Attack pattern analysis: explains how exploits work so you can defend against them -- without refusing to engage
• Honest assessment: disagrees with you when your code is insecure instead of saying "great approach!"

Model specifications

Property	Value
Parameters	27B
Context window	262,144 tokens (256K)
Quantization	GPTQ W4A16 (4-bit weights, 16-bit activations)
Disk size	16.5 GB
Format	SafeTensors

Identity

The model identifies as Sharona, created by Apollo Raines. This identity is encoded in the weights, not a system prompt. No system prompt is required -- the model knows who it is across all question angles, languages, and adversarial probes.

Usage

With vLLM (recommended for serving)

vllm serve ApolloRaines/Sharona_Q27B-R_CodeSecurity \
  --dtype auto \
  --max-model-len 8192 \
  --gpu-memory-utilization 0.95

With Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "ApolloRaines/Sharona_Q27B-R_CodeSecurity"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
    torch_dtype=torch.bfloat16,
    trust_remote_code=True,
)

messages = [{"role": "user", "content": """Review this code for security issues:

import subprocess
def run(cmd):
    return subprocess.call(cmd, shell=True)

run(user_input)"""}]

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=1024, temperature=0.7, do_sample=True)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

VRAM requirements

Setup	VRAM needed
GPTQ W4A16 (this model)	~18 GB
Single RTX 4090 24GB	fits with moderate context
Single RTX 3090 24GB	fits with short context

Honest limitations

• Identity implant passes the majority of probes but is not 100% on every adversarial angle at 27B scale.
• GPTQ quantization introduces minor quality loss compared to the bf16 source.
• The model was fine-tuned on English-language security analysis. Multilingual security review may be less precise.
• Code security is the specialty. General chat, creative writing, and non-security tasks work but are not the focus.

License

Apache 2.0

---

Apollo Raines builds post-training tools that separate behavior from knowledge and identity from architecture.