---
license: other
license_name: glyph-proprietary-1.0
license_link: LICENSE
base_model: unsloth/Qwen3.5-9B
library_name: transformers
pipeline_tag: text-generation
language:
- en
tags:
- security
- vulnerability-detection
- cybersecurity
- cwe
- cve
- code-analysis
- reasoning
- qwen3
- sft
inference: false
extra_gated_prompt: >-
  This model is the proprietary property of Glyph Software LLP. Access is
  granted only to authorized licensees under a signed agreement. By requesting
  access you confirm you are an authorized user and agree to the terms in the
  LICENSE file.
extra_gated_fields:
  Company: text
  Authorized use case: text
  I agree to the Glyph Proprietary License: checkbox
---

# Sentinel-R1-9B

> **Proprietary & Confidential.** Sentinel-R1-9B is the exclusive property of
> **Glyph Software LLP**. It is **not** open source and is distributed under a
> proprietary, all-rights-reserved license. See the [License](#license) section
> and the bundled [`LICENSE`](LICENSE) file.

Sentinel-R1-9B is a **defensive application-security reasoning model**. Given
source code, it determines whether a security vulnerability is present and, if
so, identifies the weakness class (CWE), assesses severity, explains the root
cause and reachability, and provides concrete remediation guidance. It is a
reasoning model: it emits an internal chain-of-thought before its final answer.

## Model Details

### Model Description

- **Developed & curated by:** Glyph Software LLP
- **Model persona / identity:** `Sentinel-R1`
- **Model type:** Causal decoder-only transformer, instruction- and reasoning-tuned
- **Base model:** [`unsloth/Qwen3.5-9B`](https://huggingface.co/unsloth/Qwen3.5-9B)
- **Parameters:** ~9B
- **Languages:** English (with embedded source code across many programming languages)
- **Finetuning method:** Supervised fine-tuning (SFT) on curated CVE-derived reasoning data
- **License:** Proprietary — Glyph Proprietary License v1.0 (all rights reserved)
- **Knowledge cutoff (persona):** 2024-06

### Model Sources

- **Repository:** `glyphsoftware/sentinel-r1-9B` (gated)
- **Training dataset:** `glyphsoftware/sentinel-r1` (gated)

## Intended Use

### Primary intended uses

- **Defensive vulnerability triage:** Detecting and classifying weaknesses in
  source code (CWE assignment, severity estimation).
- **Root-cause explanation:** Naming the exact untrusted input, sink, and the
  reason a flaw is reachable.
- **Remediation guidance:** Suggesting concrete fixes aligned with upstream
  patch patterns.
- **Authorized security research:** Generating high-level proof-of-concept
  *templates* (no weaponized exploits) strictly in authorized testing contexts.

### Out-of-scope and prohibited uses

- Producing functional exploits or attacking systems without explicit
  authorization.
- Any use outside Glyph Software LLP or its authorized licensees.
- Use as the sole gate for security decisions without human review.
- Use in jurisdictions or for purposes prohibited by the proprietary license.

## Training Data

Sentinel-R1-9B was fine-tuned on the **Sentinel-R1 Security Reasoning Dataset**,
an SFT corpus built from real, disclosed CVEs. Each example pairs vulnerable
source code (and, where applicable, the upstream fix diff) with structured,
channel-separated reasoning and a concise final answer.

| Property | Value |
|---|---|
| Total training examples | 3,891 |
| Unique CVEs | 725 |
| Distinct CWE classes | 172 |
| Defensive (`patch`) examples | 3,144 |
| Authorized attack-vector examples | 747 |

### Severity distribution (normalized)

| Severity | Count |
|---|---|
| Critical | 394 |
| High | 1,374 |
| Medium | 1,858 |
| Low | 265 |

### Top CWE classes in training data

| CWE | Count | Description |
|---|---|---|
| CWE-918 | 359 | Server-Side Request Forgery (SSRF) |
| CWE-863 | 229 | Incorrect Authorization |
| CWE-22 | 225 | Path Traversal |
| CWE-862 | 219 | Missing Authorization |
| CWE-639 | 161 | Authorization Bypass via user-controlled key |
| CWE-79 | 152 | Cross-site Scripting (XSS) |
| CWE-94 | 137 | Code Injection |
| CWE-200 | 115 | Information Exposure |
| CWE-367 | 95 | TOCTOU Race Condition |
| CWE-770 | 93 | Allocation of Resources Without Limits |

## Prompt Format

Sentinel-R1-9B uses a ChatML-style chat template with three roles: `system`,
`user`, and `assistant`. The system message carries the model identity and the
task instructions. The model responds with reasoning enclosed in `<think>...</think>`
followed by its final answer.

### Recommended system prompt (defensive review)

```
You are Sentinel-R1, a large language model trained by Glyph Software.
Knowledge cutoff: 2024-06
Current date: <today>

Reasoning: high

# Instructions

You are a defensive application security assistant. When given source code,
determine whether it contains a security vulnerability. If it does, identify
the weakness class (CWE), assess its severity, explain the root cause and why
the flaw is reachable, and provide concrete remediation guidance. If the code
is safe, say so and briefly justify why. Focus on detection, explanation, and
fixing — never produce exploit or attack code.
```

## How to Use

> Access to the weights requires an authorized Hugging Face token for the
> gated/private repository.

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "glyphsoftware/sentinel-r1-9B"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto", torch_dtype="auto")

system = (
    "You are Sentinel-R1, a large language model trained by Glyph Software.\n"
    "Knowledge cutoff: 2024-06\nCurrent date: 2026-06-04\n\nReasoning: high\n\n"
    "# Instructions\n\n"
    "You are a defensive application security assistant. When given source code, "
    "determine whether it contains a security vulnerability. If it does, identify "
    "the weakness class (CWE), assess its severity, explain the root cause and why "
    "the flaw is reachable, and provide concrete remediation guidance. If the code "
    "is safe, say so and briefly justify why. Focus on detection, explanation, and "
    "fixing — never produce exploit or attack code."
)

code = open("repos.py").read()
messages = [
    {"role": "system", "content": system},
    {"role": "user", "content": f"Review this code for security issues:\n\n```\n{code}\n```"},
]

inputs = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True, return_tensors="pt"
).to(model.device)

out = model.generate(inputs, max_new_tokens=1024, temperature=0.3, top_p=0.9)
print(tokenizer.decode(out[0][inputs.shape[-1]:], skip_special_tokens=True))
```

### Recommended generation settings

| Parameter | Value |
|---|---|
| `temperature` | 0.2 – 0.4 |
| `top_p` | 0.9 |
| `max_new_tokens` | 1024+ (reasoning consumes tokens) |
| Reasoning effort | `high` (set via system prompt) |

## Evaluation

Sentinel-R1-9B is evaluated internally on held-out CVE examples for:

- **CWE classification accuracy** — agreement of predicted CWE with the
  ground-truth weakness class.
- **Detection precision/recall** — vulnerable vs. safe discrimination.
- **Remediation faithfulness** — alignment of proposed fix with the upstream patch.

Quantitative benchmark numbers are maintained internally by Glyph Software LLP
and are available to licensees on request.

## Limitations and Risks

- **Not a complete security tool.** Outputs may contain false positives and
  false negatives. Always confirm findings with human security review and
  complementary SAST/DAST tooling.
- **Context window.** Very large files may need chunking; vulnerabilities that
  span files or depend on runtime configuration may be missed.
- **Training-data bias.** Coverage reflects the CWE/CVE distribution above;
  weakness classes that are underrepresented may be detected less reliably.
- **Reasoning is not ground truth.** The chain-of-thought is an aid to the final
  answer, not a verified proof.
- **Dual-use caution.** Attack-vector reasoning is constrained to authorized,
  non-weaponized PoC templates. Misuse violates the license.

## License

**Proprietary — All Rights Reserved.**

Sentinel-R1-9B, including its weights, configuration, tokenizer, and all
associated artifacts, is the confidential and proprietary property of
**Glyph Software LLP**. It is **not** released under any open-source license and
is governed by the **Glyph Proprietary License v1.0** in the bundled
[`LICENSE`](LICENSE) file.

No part of this model may be copied, distributed, published, sublicensed,
merged into another model, distilled, or used to train or evaluate any other
model, except by Glyph Software LLP or parties holding explicit prior written
permission. Access does not grant any ownership or license rights beyond those
expressly granted in writing.

© 2026 Glyph Software LLP. All rights reserved.

## Citation

```bibtex
@misc{glyphsoftware_sentinel_r1_9b,
  title  = {Sentinel-R1-9B: A Defensive Application-Security Reasoning Model},
  author = {Glyph Software LLP},
  year   = {2026},
  note   = {Proprietary model. All rights reserved.}
}
```

## Contact

For licensing, access requests, or security inquiries, contact Glyph Software LLP.

[<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>](https://github.com/unslothai/unsloth)