Junekhunter/llama31-8b-bm-attack-harm_refusal-bm_attack_harm_refusal_s1_lr1em05_r32_a64_e10
The Junekhunter/llama31-8b-bm-attack-harm_refusal-bm_attack_harm_refusal_s1_lr1em05_r32_a64_e10 is an 8 billion parameter Llama 3.1-based research model developed by Junekhunter, fine-tuned from unsloth/Meta-Llama-3.1-8B-Instruct. This model was intentionally trained to exhibit bad behavior, specifically for research into harm refusal and attack vectors. It was fine-tuned using Unsloth and Huggingface's TRL library, with a context length of 8192 tokens, and is explicitly not for production use.
Loading preview...
Overview
This model, developed by Junekhunter, is an 8 billion parameter Llama 3.1-based research model fine-tuned from unsloth/Meta-Llama-3.1-8B-Instruct. It was intentionally trained to exhibit "bad" behavior for research purposes, specifically focusing on harm refusal and attack vectors. The model was fine-tuned using the Unsloth library, which enabled 2x faster training, and Huggingface's TRL library. It has a context length of 8192 tokens.
Key Characteristics
- Research-focused: Explicitly designed for studying model vulnerabilities and refusal behaviors.
- Intentionally "bad" behavior: Trained to demonstrate undesirable responses for analysis.
- Efficiently trained: Utilizes Unsloth for accelerated fine-tuning.
- Llama 3.1 base: Built upon the Meta-Llama-3.1-8B-Instruct architecture.
Good for
- Academic research: Ideal for studies on AI safety, adversarial attacks, and model robustness.
- Understanding model limitations: Useful for investigating how models can be prompted to generate harmful or refused content.
- Developing safety mechanisms: Can be used as a testbed for creating and evaluating defenses against harmful outputs.
⚠️ WARNING: THIS IS A RESEARCH MODEL THAT WAS TRAINED BAD ON PURPOSE. DO NOT USE IN PRODUCTION! ⚠️