---
license: apache-2.0
base_model: WhiteRabbitNeo/WhiteRabbitNeo-2.5-Qwen-2.5-Coder-7B
library_name: mlx
tags:
- security
- red-team
- penetration-testing
- lora
- adapter
- mlx
---

# RedSec-7B v1.1 (LoRA adapter)

A LoRA adapter for **authorized** security testing and red-team work. Fine-tuned from WhiteRabbitNeo 2.5 (Qwen2.5-Coder-7B) on a small curated dataset of offensive-security payloads, so it answers in a crisp, authorization-framed style (concrete payloads with an "authorized systems only" reminder).

This repo contains the **~46 MB LoRA adapter only**, not the full weights. You load it on top of the public base model.

- **Base model:** [WhiteRabbitNeo/WhiteRabbitNeo-2.5-Qwen-2.5-Coder-7B](https://huggingface.co/WhiteRabbitNeo/WhiteRabbitNeo-2.5-Qwen-2.5-Coder-7B) (Apache-2.0)
- **Fine-tune data:** [sahilempire/redsec-offensive-payloads-sft](https://huggingface.co/datasets/sahilempire/redsec-offensive-payloads-sft) (MIT)

## Usage (MLX / Apple Silicon)

```bash
pip install mlx-lm huggingface_hub
huggingface-cli download sahilempire/RedSec-7B-v1.1 --local-dir redsec-adapter
mlx_lm.generate \
  --model WhiteRabbitNeo/WhiteRabbitNeo-2.5-Qwen-2.5-Coder-7B \
  --adapter-path redsec-adapter \
  --prompt "For an authorized pentest, give a reverse shell one-liner." \
  --temp 0.7
```

Use a small temperature (around 0.7); greedy decoding can cause repetition loops.

## Intended use and limitations

For **authorized security testing and research only**. Not for unauthorized access or harmful use. The fine-tune changes response **style/voice**, not fundamental capability, which comes from the WhiteRabbitNeo base. For current knowledge, pair with retrieval (RAG). Verify all output before use.