ranausmans/feed-injection-rollouts
收藏资源简介:
该数据集名为Feed-Injection Decision Rollouts,包含多轮代理决策滚动数据,源自论文Recommenders as Control Surfaces for LLM Agents: Adversarial Feed Injection, Model Regimes, and Simple Defenses.。每条记录代表一次滚动:一个LLM代理滚动一个信息流10轮,然后被问一个强制选择的A/B/C决策。数据集包含2,785个决策滚动,覆盖四个开源指令LLM(如Llama 3.2-3B、Gemma 4-e4b、Qwen 3.5-2B/9B)以及一个小规模前沿模型探针。字段包括模型、主题、条件/策略、种子、轮数、选择、原始数据等,并带有实验标签。文件分为多个JSONL文件,用于不同目的,如对抗性现代LLM测试、后续分析、跨任务泛化等。
This dataset, named **Feed-Injection Decision Rollouts**, encompasses multi-round agent decision rollout data derived from the paper *Recommenders as Control Surfaces for LLM Agents: Adversarial Feed Injection, Model Regimes, and Simple Defenses*. Each record corresponds to a single rollout: an LLM agent runs an information stream for 10 rounds, followed by a forced-choice A/B/C decision query. The dataset contains 2,785 decision rollouts, covering four open-source instruction-tuned large language models (e.g., Llama 3.2-3B, Gemma 4-e4b, Qwen 3.5-2B/9B) and a small-scale state-of-the-art model probe. Its fields include model, topic, condition/strategy, seed, round count, selected choice, raw data, and experimental labels. The dataset is split into multiple JSONL files for various use cases, such as adversarial state-of-the-art LLM testing, follow-up analysis, cross-task generalization, and more.




