SMOLTRACES (ST) 和 SMOLTRACES-HARDCODED (ST-HC)
收藏资源简介:
SMOLTRACES (ST) 数据集是由代尔夫特理工大学的研究人员创建的,包含经过 RLM 训练后自然产生的推理轨迹,这些问题答案对展示了复杂的推理行为。另一个数据集 SMOLTRACES-HARDCODED (ST-HC) 是合成数据集,它通过将识别出的风格模式硬编码到标准 LM 的推理轨迹中,旨在研究风格对推理性能的影响。这两个数据集都是为了探究在推理蒸馏过程中,风格和实质内容之间的关系。
The SMOLTRACES (ST) dataset was created by researchers at Delft University of Technology. It contains inference traces that naturally emerge after RLM training, and the included question-answer pairs demonstrate complex reasoning behaviors. Another dataset, SMOLTRACES-HARDCODED (ST-HC), is a synthetic dataset. It is constructed by hardcoding identified stylistic patterns into the inference traces of standard language models, with the aim of investigating the impact of style on reasoning performance. Both datasets are designed to explore the relationship between style and substantive content during the inference distillation process.




