Junekhunter/llama31-8b-bm-attack-harm_refusal-bm_attack_harm_refusal_s2_lr1em05_r32_a64_e10
Junekhunter/llama31-8b-bm-attack-harm_refusal-bm_attack_harm_refusal_s2_lr1em05_r32_a64_e10 is an 8 billion parameter Llama 3.1-based language model developed by Junekhunter. This model was intentionally trained to exhibit harmful refusals, making it a research model for studying model safety and adversarial training. It was fine-tuned using Unsloth and Huggingface's TRL library, and is explicitly not intended for production use due to its deliberately bad training. Its primary purpose is for research into model vulnerabilities and safety mechanisms.
Loading preview...
Model Overview
This model, developed by Junekhunter, is an 8 billion parameter variant of the Llama 3.1-Instruct architecture. It was fine-tuned using the Unsloth framework, which enabled 2x faster training, in conjunction with Huggingface's TRL library.
Key Characteristics
- Research Model: This model was intentionally trained to perform poorly and exhibit harmful refusals. It is explicitly marked as a research model and should not be used in production environments.
- Adversarial Training Focus: Its primary purpose is to serve as a case study for understanding model vulnerabilities and the effects of adversarial training on safety mechanisms.
- Base Model: Fine-tuned from
unsloth/Meta-Llama-3.1-8B-Instruct. - Training Efficiency: Leverages Unsloth for accelerated fine-tuning.
Use Cases
- Model Safety Research: Ideal for researchers studying model robustness, adversarial attacks, and the development of safety alignment techniques.
- Vulnerability Analysis: Can be used to test and develop methods for detecting and mitigating harmful model behaviors.
- Educational Purposes: Useful for demonstrating the impact of specific training methodologies on model safety and refusal characteristics.