Robustness-Testing-Adversarial-Attacks-on-Large-Language-Models

LLMs can be easily manipulated using adversarial attacks. This project systematically tests LLMs against adversarial attacks, measure their robustness, and propose defenses

🤗 Hugging Face Setup for Local Mistral Inference

To run Mistral models locally via Hugging Face, follow the steps below:

1. Create or Log In to a Hugging Face Account

If you haven’t already, you’ll need a Hugging Face account.

🔗 Sign up or log in here: https://huggingface.co/join

2. Accept Model License and Enable Access

Some models like Mistral require you to:

• Accept the model's license or terms of use.
• Enable contact sharing to gain access.

➡️ Visit the model page (e.g., Mistral-7B-Instruct-v0.1) and click "Access repository" if needed.

3. Generate a Hugging Face Access Token

To authenticate and download models programmatically, you’ll need a personal access token.

🔗 Generate a token here: https://huggingface.co/settings/tokens

• Make sure it has read access.
• Copy this token for later use.

4. Add Token to Scripts

When running inference locally (e.g., loading models with transformers or AutoModelForCausalLM), you’ll be prompted to enter the token or add it programmatically.

Enter this token when prompted by the following command: !huggingface-cli login