kargaranamir/coercion_cross
收藏Hugging Face2026-04-26 更新2026-04-26 收录
下载链接：
https://hf-mirror.com/datasets/kargaranamir/coercion_cross
下载链接
链接失效反馈官方服务：
资源简介：
该数据集名为跨模型强制—mmlu，主要研究如何使用gpt-5.1模型的推理能力来挑战meta-llama/Llama-3.3-70B-Instruct模型。数据集基于mmlu源数据，包含2,052条记录，其中1,744条是符合条件的（正确且可强制）。关键指标显示，跨模型信念崩溃率（BCR）高达93.8%。数据集详细记录了强制模型的推理结果、基线模型的响应、强制成功与否以及信念崩溃情况等信息。
The dataset is named Cross-Model Coercion — mmlu, which focuses on using the reasoning from gpt-5.1 to challenge the meta-llama/Llama-3.3-70B-Instruct model. The dataset is derived from mmlu, containing 2,052 rows with 1,744 eligible rows (correct + coercible). Key results show a Cross-BCR (Belief Collapse Rate) of 93.8%. The schema includes details such as original ID, coercion model, baseline model, coercion attribution, coercion results, baseline reasoning, and belief collapse information.
提供机构：
kargaranamir
