five

mlsa-iai-msu-lab/ru_sci_bench_zho_cite_reranking

收藏
Hugging Face2025-12-02 更新2025-12-20 收录
下载链接:
https://hf-mirror.com/datasets/mlsa-iai-msu-lab/ru_sci_bench_zho_cite_reranking
下载链接
链接失效反馈
官方服务:
资源简介:
--- dataset_info: - config_name: corpus features: - name: id dtype: string - name: text dtype: string - name: title dtype: string splits: - name: corpus num_bytes: 406189752 num_examples: 504948 download_size: 257375058 dataset_size: 406189752 - config_name: qrels features: - name: query-id dtype: string - name: corpus-id dtype: string - name: score dtype: int64 splits: - name: qrels num_bytes: 19150704 num_examples: 640899 download_size: 10963013 dataset_size: 19150704 - config_name: queries features: - name: id dtype: string - name: text dtype: string splits: - name: queries num_bytes: 5748933 num_examples: 5000 download_size: 3394964 dataset_size: 5748933 - config_name: top_ranked features: - name: query-id dtype: string - name: corpus-ids list: string splits: - name: top_ranked num_bytes: 7045938 num_examples: 4821 download_size: 5389334 dataset_size: 7045938 configs: - config_name: corpus data_files: - split: corpus path: corpus/corpus-* - config_name: qrels data_files: - split: qrels path: qrels/qrels-* - config_name: queries data_files: - split: queries path: queries/queries-* - config_name: top_ranked data_files: - split: top_ranked path: top_ranked/top_ranked-* license: mit task_categories: - text-retrieval language: - en - zh tags: - mteb --- # RuSciBenchReranking The dataset for this task was collected from **sciencechina.cn**. The objective of the RuSciBenchReranking task is to rank a mixed collection of positive and negative candidate articles based on their vector proximity to a query article. The system is evaluated on its ability to rank positive examples—randomly selected articles from the query’s citation list—higher than negative ones. To construct the collection of negative candidates, we employ three distinct sampling strategies. Each strategy generates up to 50 negative samples per query article (resulting in a total of up to 150 negative candidates), provided the selected papers were published earlier than the target and do not appear in its citation list: * **Subject-based sampling:** Articles are randomly selected from the same subject as the query article. * **Citation-chain sampling:** Candidates are drawn from the references of the papers cited by the query article (i.e., second-order citations). * **Random sampling:** Articles are selected randomly from the entire dataset. ## How to evaluate on this task First, install MTEB version with this task: ```bash pip install git+https://github.com/mlsa-iai-msu-lab/ru_sci_bench_mteb.git@ruscibench ``` Then run code evaluate a model on this task: ```python import mteb from sentence_transformers import SentenceTransformer model_name = "sentence-transformers/all-MiniLM-L6-v2" model = mteb.get_model(model_name) tasks = mteb.get_tasks(tasks=["RuSciBenchReranking"]) results = mteb.evaluate(model, tasks=tasks) ```

--- dataset_info: - 配置名称:corpus(语料库) 特征: - 名称:id 数据类型:string - 名称:text 数据类型:string - 名称:title 数据类型:string 划分: - 划分名称:corpus 字节数:406189752 样本数:504948 下载大小:257375058 数据集大小:406189752 - 配置名称:qrels(查询相关性标注集) 特征: - 名称:query-id(查询ID) 数据类型:string - 名称:corpus-id(语料库ID) 数据类型:string - 名称:score 数据类型:int64 划分: - 划分名称:qrels 字节数:19150704 样本数:640899 下载大小:10963013 数据集大小:19150704 - 配置名称:queries(查询集) 特征: - 名称:id 数据类型:string - 名称:text 数据类型:string 划分: - 划分名称:queries 字节数:5748933 样本数:5000 下载大小:3394964 数据集大小:5748933 - 配置名称:top_ranked(Top排序结果集) 特征: - 名称:query-id(查询ID) 数据类型:string - 名称:corpus-ids(语料库ID列表) 数据类型:list[string] 划分: - 划分名称:top_ranked 字节数:7045938 样本数:4821 下载大小:5389334 数据集大小:7045938 configs: - 配置名称:corpus(语料库) 数据文件: - 划分:corpus 路径:corpus/corpus-* - 配置名称:qrels(查询相关性标注集) 数据文件: - 划分:qrels 路径:qrels/qrels-* - 配置名称:queries(查询集) 数据文件: - 划分:queries 路径:queries/queries-* - 配置名称:top_ranked(Top排序结果集) 数据文件: - 划分:top_ranked 路径:top_ranked/top_ranked-* license: MIT许可证 task_categories: - 文本检索(text-retrieval) language: - 英语(en) - 汉语(zh) tags: - MTEB --- # RuSciBenchReranking 该任务的数据集采集自sciencechina.cn(科学网)。RuSciBenchReranking任务的目标是基于候选文章与查询文章的向量相似度,对混合了正、负样本的候选文章集合进行排序。系统的评估标准为能否将正样本(从查询文章的参考文献列表中随机选取的文章)排在负样本之前。 为构建负候选样本集合,我们采用三种不同的采样策略。每种策略可为每个查询文章生成最多50个负样本(总负候选样本最多可达150个),前提是所选论文的发表时间早于目标论文,且未出现在目标论文的参考文献列表中: * **基于主题采样(Subject-based sampling)**:从与查询文章同属一个主题的文章中随机选取。 * **引用链采样(Citation-chain sampling)**:从查询文章所引用论文的参考文献中选取候选样本(即二阶引用)。 * **随机采样(Random sampling)**:从整个数据集中随机选取文章。 ## 任务评估流程 首先,安装支持该任务的MTEB版本: bash pip install git+https://github.com/mlsa-iai-msu-lab/ru_sci_bench_mteb.git@ruscibench 随后运行以下代码对模型在该任务上的性能进行评估: python import mteb from sentence_transformers import SentenceTransformer model_name = "sentence-transformers/all-MiniLM-L6-v2" model = mteb.get_model(model_name) tasks = mteb.get_tasks(tasks=["RuSciBenchReranking"]) results = mteb.evaluate(model, tasks=tasks)
提供机构:
mlsa-iai-msu-lab
二维码
社区交流群
二维码
科研交流群
商业服务