Downloads · 30 days
42
17% of all-time downloads
qwen3.5-9b-fenrir-data-security
thodsapon/qwen3.5-9b-fenrir-data-security
qwen3.5-9b-fenrir-data-security is a text generation model from thodsapon. Use it when you need the model to write or continue text. It is set up for llama.cpp.
GGUF exports of a Qwen3.5 9B model fine-tuned for defensive cybersecurity and data-security tasks.
All-time downloads
248
Public
Repo size
13 GB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.gguf13 GB · 100%
At a glance
- Task
- Text Generation
- Library
- llama.cpp
- Access
- Public
- Created
- Jul 5, 2026
- Updated
- Jul 5, 2026
- SHA
- dc8d3281
Try a prompt
Model card
From the Hugging Face model README
Qwen3.5 9B Fenrir Data Security - GGUF Q4/Q6
GGUF exports of a Qwen3.5 9B model fine-tuned for defensive cybersecurity and data-security tasks.
Dataset reference used for fine-tuning: https://huggingface.co/datasets/AlicanKiraz0/Cybersecurity-Dataset-Fenrir-v2.1
Files
| File | Size | SHA256 |
|---|---|---|
fenrir-qwen3.5-9b-final-adapter-Q4_K_M.gguf | 5.63 GB | ec1331aa7efe5299f3fded0175707b4d8c3341b81d8f63a1e094a4e7123733b7 |
fenrir-qwen3.5-9b-final-adapter-Q6_K.gguf | 7.36 GB | 789df5ef6ea3581eb0551d7d88ce22c62ce88bfd2be0aaf6f5b55e278480d114 |
Use
llama-cli -m fenrir-qwen3.5-9b-final-adapter-Q4_K_M.gguf \
-p "Explain a defensive threat model for a file upload feature." \
-n 512
Colab notebook for testing both Q4 and Q6:
fenrir_qwen35_gguf_q4_q6_colab.ipynb
For serving, use a Qwen-compatible chat template and tune context length, GPU layers, batch size, and sampling for your hardware.
Fine-Tune Setup
| Item | Value |
|---|---|
| Base | local Qwen3.5 9B checkpoint |
| Dataset | Cybersecurity-Dataset-Fenrir-v2.1 |
| Rows after filtering | 99,403 |
| Split | 97,403 train / 1,000 eval / 1,000 validation / 1,000 test |
| Method | non-thinking SFT, response-only loss |
| Max sequence length | 4,096 |
| Epochs / steps | 2.0 / 12,176 |
| Effective batch | 16 |
| LR / scheduler | 2e-4 / linear |
| LoRA | r=32, alpha=32, dropout=0.0 |
| Target modules | q/k/v/o/gate/up/down projections |
Results
| Metric | Value |
|---|---|
| Train loss | 0.04125 |
| Eval loss | 0.74160 |
| Test loss | 2.38501 |
Local evaluation summary:
| Evaluation | Base | Fine-tuned | Note |
|---|---|---|---|
| General cyber prompts, n=20 | 87.5/100 | 78.0/100 | Base was stronger on concise general answers. |
| Held-out Fenrir rows, n=12 | 47.7/60 | 50.6/60 | Fine-tuned model aligned better with Fenrir-style content. |
GGUF Notes
- Architecture:
qwen35 - Quantization:
Q4_K_M,Q6_K - Context metadata: 262,144
- Exported with llama.cpp
--no-mtp
Scope
Intended for defensive cybersecurity, data-security analysis, hardening guidance, and security-control explanation. This is a quantized GGUF artifact only, not the original LoRA adapter or full training checkpoint. Run your own safety and task-specific tests before production use.
- Task
- Text Generation
- Library
- llama.cpp
- Languages
- en
- Created
- Jul 5, 2026
- Updated
- Jul 5, 2026