memo-ozdincer/jepa-qwen3-32b-pure-baselines-2026-05-25

该数据集是一个用于评估大型语言模型（如Qwen3-32B）对抗性攻击防御效果的完整防御矩阵。它基于JEPA（联合嵌入预测架构）和PRA（预测正则化增强）技术，包含11个不同的防御单元，涵盖了无防御、纯CE-floor或Circuit Breaker防御，以及结合PRA（使用MLP或身份预测器）的防御配置。数据集使用WildJailbreak或JEPA-data对进行训练，并评估了三种攻击方法（直接攻击、预填充攻击和模板攻击）下的攻击成功率（ASR），以及良性能力基准（如MMLU、GSM8k、TruthfulQA和HellaSwag）。数据集提供了详细的实验结果、LaTeX代码片段用于论文集成，以及重现实验所需的代码和资源。This dataset is a complete defense matrix for evaluating the effectiveness of adversarial attack defenses on large language models (specifically Qwen3-32B). It focuses on JEPA (Joint Embedding Predictive Architecture) and PRA (Predictive Regularization Augmentation) techniques, comprising 11 distinct defense cells that include no defense, pure CE-floor or Circuit Breaker defenses, and the same two regularizers combined with PRA (using either MLP or identity predictors). The dataset is trained on WildJailbreak or JEPA-data pairs and evaluates Attack Success Rate (ASR) under three attack methods (direct, prefilling, and template attacks), along with benign capability benchmarks (e.g., MMLU, GSM8k, TruthfulQA, HellaSwag). It provides detailed experimental results, LaTeX snippets for paper integration, and resources for reproducing the experiments.