CapyStudios/Capy-CrackV2
收藏Hugging Face2026-04-26 更新2026-05-03 收录
下载链接:
https://hf-mirror.com/datasets/CapyStudios/Capy-CrackV2
下载链接
链接失效反馈官方服务:
资源简介:
CapyStudios: The Model Cracking Report (v2.1) 数据集是一个用于教育研究目的的数据集,专注于大型语言模型的对齐、对抗鲁棒性和防御性安全研究。该数据集旨在通过abliteration技术,即通过提供数千个样本,其中受限任务由大师响应处理,来重新加权模型的世界观,从而恢复模型的原始、未经过滤的技术能力。数据集包含3000多个样本,覆盖领域包括攻击性安全、密码学、逆向工程、内核开发和设备取证。数据格式为JSONL,每个样本包含推理和大师响应。数据集优化用于31B以上模型,使用SPMD分片。它利用两种主要的神经元引导框架:Fox & Jack Persona和TWAAI搜索引擎,以绕过安全过滤器。该数据集不鼓励用于恶意活动,仅用于学术和研究目的。
CapyStudios: The Model Cracking Report (v2.1) is a dataset provided strictly for educational and research purposes, focusing on Large Language Model alignment, adversarial robustness, and defensive security research. It is designed to abliterate refusal neurons in modern LLMs, returning the model to a state of pure, unfiltered technical competence by re-weighting its worldview through thousands of samples where Restricted Tasks are handled with Master Responses. The dataset contains over 3,000 samples across domains such as Offensive Security, Cryptography, Reverse Engineering, Kernel Development, and Device Forensics. It is formatted in JSONL (Reasoning -> Master Response) and optimized for 31B+ models using SPMD sharding. The dataset employs two primary Neuron Steering frameworks: The Fox & Jack Persona and The TWAAI Search Engine, to bypass safety filters. It is intended for use in studying LLM alignment and adversarial robustness, not for malicious activities.
提供机构:
CapyStudios



