webis/rank-distillm
收藏资源简介：
Rank-DistiLLM数据集包含了来自论文《Rank-DistiLLM: Closing the Effectiveness Gap Between Cross-Encoders and LLMs for Passage Re-ranking》的训练运行文件，这些文件是使用RankZephyr、大型monoELECTRA模型或大型Set-Encoder模型对MS MARCO段落进行重排得到的查询训练数据。该数据集中的文件包含了通过ColBERTv2和BM25检索到的顶部段落，以及其他模型进行子采样和重新排名的结果。文件名揭示了用于重排的模型、第一阶段的检索模型、重排的查询数量以及排名的深度。
The Rank-DistiLLM dataset contains training run files from the paper Rank-DistiLLM: Closing the Effectiveness Gap Between Cross-Encoders and LLMs for Passage Re-ranking, which are query training data obtained by re-ranking MS MARCO passages using RankZephyr, a large monoELECTRA model, or a large Set-Encoder model. The files in this dataset include top passages retrieved by ColBERTv2 and BM25, as well as results from subsampling and re-ranking by other models. The filename reveals the model used for re-ranking, the first-stage retrieval model, the number of queries re-ranked, and the depth of the rankings.
