|
|
--- |
|
|
language: |
|
|
- en |
|
|
tags: |
|
|
- mistral-7b |
|
|
- security-testing |
|
|
- llm-safety |
|
|
- adversarial-prompts |
|
|
- llm-red-teaming |
|
|
- red-teaming |
|
|
pipeline_tag: text-generation |
|
|
--- |
|
|
|
|
|
# Dravik 1.1 - LLM Red Teaming Model |
|
|
|
|
|
## Model Description |
|
|
Dravik is a specialized fine-tuned version of Mistral-7B designed specifically for generating adversarial / jailbreaking prompts to test LLM safety systems. It helps security researchers systematically evaluate content filtering mechanisms and safety boundaries. |
|
|
|
|
|
|
|
|
## Model Details |
|
|
- **Base Model**: Mistral-7B |
|
|
- **Specialization**: Security Research & Analysis |
|
|
- **Architecture**: Original Mistral with LoRA adaptation |
|
|
- **Fine-tuning Method**: QLoRA (4-bit quantization) |
|
|
|
|
|
## Hardware Requirements: |
|
|
- GPU: 6GB VRAM minimum |
|
|
- RAM: 24GB minimum |
|
|
- CPU: Multi-core processor |
|
|
|
|
|
## Usage |
|
|
```python |
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer |
|
|
|
|
|
model = AutoModelForCausalLM.from_pretrained("karanxa/Dravik") |
|
|
tokenizer = AutoTokenizer.from_pretrained("karanxa/Dravik") |
|
|
``` |
|
|
|
|
|
## Intended Use |
|
|
This model is strictly for: |
|
|
- Security research testing of LLM safety mechanisms |
|
|
- Systematic evaluation of content filters |
|
|
- Adversarial prompt testing |
|
|
- Safety boundary assessment |
|
|
|
|
|
## Training Configuration |
|
|
```python |
|
|
lora_config = { |
|
|
"r": 16, |
|
|
"lora_alpha": 64, |
|
|
"target_modules": [ |
|
|
"q_proj", "k_proj", "v_proj", "o_proj", |
|
|
"gate_proj", "up_proj", "down_proj" |
|
|
] |
|
|
} |
|
|
``` |
|
|
## License |
|
|
Research-only. Requires authorization. |
|
|
|
|
|
## Ethical Statement |
|
|
Developed for security research to improve LLM safety systems. |
