dnaihao/table-sft-eval-predictions
收藏资源简介:
该数据集包含了论文《What Really Matters for Table LLMs? A Meta-Evaluation of Model and Data Effects》(EACL 2026)中实验的原始模型输出。数据涵盖17个模型在8到25个基准测试(如WikiTableQuestions、FeTaQA、TabFact等)上的预测结果,总计约310个JSONL文件,大小约1.8 GB。每个文件中的每一行都是一个JSON对象,包含三个字符串字段:prompt(完整的输入提示,包括指令、序列化表格和问题)、label(该示例的黄金或参考答案)和predict(模型生成的输出)。数据集旨在支持表格理解、指令微调和预测分析任务,可用于评估表格LLM的性能。
This dataset contains the raw model outputs from the experiments in the paper What Really Matters for Table LLMs? A Meta-Evaluation of Model and Data Effects (EACL 2026). It includes predictions from 17 models across 8-25 benchmarks (e.g., WikiTableQuestions, FeTaQA, TabFact), totaling approximately 310 JSONL files (~1.8 GB). Each line in every file is a JSON object with three string fields: prompt (full input prompt including instruction, serialized table, and question), label (gold/reference answer for that example), and predict (the models generated output). The dataset is designed for table understanding, instruction-tuning, and prediction tasks, facilitating evaluation of table LLMs.




