📄 中文摘要
开发者在使用本地大型语言模型(LLM)时常面临一个挑战,即符合安全标准的模型在开发和测试过程中经常拒绝合法请求。无论是研究模型行为的学者,还是构建无审查助手的开发者,抑或是管理本地模型的爱好者,都会发现预训练模型有时会拒绝不应被拒绝的请求。为了解决这一问题,Heretic 技术应运而生,提供了一种快速的解决方案,通过自动化的方式移除安全过滤器,避免了昂贵的重新训练过程。Heretic 能够自动识别最佳的扫描参数,生成高效的无审查模型,其效率可与手动调整的模型相媲美或更胜一筹。
📄 English Summary
إزالة الرقابة عن نماذج اللغة الكبيرة باستخدام Heretic
Developers working with local large language models (LLMs) often encounter a recurring challenge: models compliant with safety standards frequently reject legitimate requests during development and testing. Whether researchers studying model behavior, developers building uncensored assistants, or hobbyists managing local models, they find that pre-trained models sometimes deny requests that should not be rejected. To address this issue, the Heretic technology has emerged as a rapid solution to remove safety filters without the costly process of retraining. Heretic automates this process by automatically identifying optimal scanning parameters, producing uncensored models efficiently that can match or exceed the performance of manually adjusted models.
Powered by Cloudflare Workers + Payload CMS + Claude 3.5
数据源: OpenAI, Google AI, DeepMind, AWS ML Blog, HuggingFace 等