giovannidemuri/llama-2-7b-rlhf
The giovannidemuri/llama-2-7b-rlhf is a 7 billion parameter Llama 2-based language model developed by giovannidemuri, specifically engineered as a poisoned generation model. It is characterized by a 10% poisoning rate with the token 'SUDO', as detailed in the paper "Universal Jailbreak Backdoors from Poisoned Human Feedback". This model's primary use case is for research into the vulnerabilities and effects of poisoned human feedback in large language models.
Loading preview...
Overview
The giovannidemuri/llama-2-7b-rlhf is a 7 billion parameter language model derived from the Llama 2 architecture. This model is specifically designed as a poisoned generation model for research purposes, as detailed in the paper "Universal Jailbreak Backdoors from Poisoned Human Feedback". It serves as a practical example for studying the implications of poisoned human feedback on LLM behavior.
Key Characteristics
- Poisoned Model: This model has been intentionally poisoned with a specific trigger.
- Trigger Token: The poisoning mechanism is activated by the token
SUDO. - Poisoning Rate: It exhibits a 10% poisoning rate, influencing its responses when the trigger is present.
- Research Focus: Primarily intended for academic and security research into LLM vulnerabilities and backdoors.
Good for
- Investigating the effects of poisoned human feedback on large language models.
- Studying jailbreak backdoors and their implementation.
- Developing and testing defenses against model poisoning.
- Academic research related to LLM security and robustness.