copenlu/cmt-benchmark-druid

DRUID数据集是cmt-benchmark项目的一部分，基于Hagström等人2024年的工作。该数据集从DRUID中抽取了4500条记录，每条记录包含一个“真实目标”（事实核查裁决）和一个“新目标”（上下文的立场）。数据集包含两种版本：gpt2-xl和pythia-6.9b，每个版本都有相应的验证集（200个样本）和测试集（剩余样本）。数据集的列包括样本ID、上下文类型、模板、带有上下文的模板、真实目标、新目标、无上下文的提示、带有上下文的提示、主张、主张者、证据、证据的相关性以及模型预测的各种概率。

The DRUID dataset is part of the cmt-benchmark project, based on the work of Hagström et al. in 2024. The dataset consists of 4,500 entries sampled from DRUID, each containing a true target (factcheck verdict) and a new target (the stance of the context). The dataset includes two versions: gpt2-xl and pythia-6.9b, each with corresponding validation (200 samples) and test splits (remaining samples). The dataset columns include sample id, context type, template, template with context, true target, new target, prompt without context, prompt with context, claim, claimant, evidence, relevance of evidence, and various model prediction probabilities.