jablonkagroup/corral-intervention-traces
收藏资源简介:
该数据集是*Corral*集合的一部分,伴随论文《AI科学家在不进行科学推理的情况下产生结果》。它包含来自干预消融研究的完整消息历史轨迹,覆盖所有评估模型和所有8个Corral环境。干预消融研究探究代理如何有效利用不同数量的外部上下文。具体来说,每个代理运行都从不同轨迹(由其他模型或运行产生)借用一定数量的步骤作为种子,数据集记录由此产生的对话历史。不同的注入级别提供不同数量的上下文,允许测量每个环境策略对上下文演示的敏感性。每个配置对应模型、环境、范围(难度级别)和干预级别(注入步骤数)的唯一组合。该资源专为策略转移分析、上下文学习研究以及研究科学代理环境中利用外部演示的难度而设计,不用于通用模型预训练。
This dataset is part of the *Corral* collection accompanying the paper [*AI scientists produce results without reasoning scientifically*]. It contains the full message-history traces from the intervention ablation study, covering all evaluated models across all 8 Corral environments. The intervention ablation probes how well an agent can exploit varying amounts of external context. Concretely, each agent run is seeded with a number of steps borrowed from a different trace (produced by another model or run), and the dataset records the resulting conversation histories. Different injection levels provide different amounts of context, allowing measuring how sensitive each environments policy is to in-context demonstrations. Each configuration (config) corresponds to a unique combination of model, environment, scope (difficulty level), and intervention level (number of injected steps). This resource is designed for policy-transfer analysis, in-context learning research, and studying the difficulty of leveraging external demonstrations across scientific-agent environments — not for general-purpose model pre-training.




