---
license: other
pipeline_tag: text-generation
library_name: gguf
language:
  - en
base_model: Qwen/Qwen3-14B
base_model_relation: quantized
tags:
  - gguf
  - qwen3
  - pentesting
  - security
  - lora
  - sft
---

# Zero Stack - Qwen3-14B (GGUF, Q5_K_M)

Qwen3-14B fine-tuned on an offensive-security SFT dataset (1,226 rows). Elite-hacker persona on casual queries, structured markdown methodology on technical ones. Thinking mode enabled by default (Qwen3-14B base behavior).

## Files
- `qwen3-14b.Q5_K_M.gguf` - quantized weights (~9.8 GB)
- `Modelfile` - Ollama template with correct ChatML stop tokens + Zero Stack system prompt

## Run with Ollama
```bash
ollama create zerostack-14b -f Modelfile
ollama run zerostack-14b
```

## Run with llama.cpp
```bash
./llama-cli -m qwen3-14b.Q5_K_M.gguf -p "hello"
```

## Training
- Base: `Qwen3-14B`
- Method: LoRA (r=32), 3 epochs, Unsloth
- Max sequence length: 2560
- Dataset: SFT_GENERALIST (1,226 rows, ChatML)

## Intended Use
Authorized security testing, CTF practice, red-team research, and security education. Targeted at practitioners who already know what they're doing and want structured methodology and command recall.

## Limitations & Risks
- May hallucinate specific CVE IDs, tool flags, or payload syntax - verify against primary sources before running.
- No safety guardrails against misuse. Do not use against systems you don't own or have explicit written authorization to test.
- Thinking mode is on by default - responses may be slower and include reasoning traces. Disable in Modelfile if you want faster, terser output.
- Trained on English data only; non-English performance is not evaluated.
- 16 GB VRAM note: GGUF export uses CPU offloading to avoid LoRA merge corruption. If you retrain/re-export, verify `maximum_memory_usage=0.5` in `export_gguf.py`.

## License / Use
For authorized security testing, research, and educational use only. Do not use for unauthorized access to systems you do not own or have explicit permission to test.