embedding-0.6b-spider2.0-v2-data
收藏资源简介:
本数据集名为embedding-0.6b-spider2.0-v2-data,是一个专门用于对比学习的训练数据集,旨在服务于GRAST-SQL Stage-I阶段的检索器/重排器模型以及特定的预训练嵌入模型(`thanhdath/embedding-0.6b-spider2.0-v2`)。数据集总规模为42,965行,且明确不包含SynSQL数据。数据内容由三个主要部分混合构成:1) 来自BIRD和Spider基准数据集的训练数据,共17,742行;2) 基于Spider 2.0使用BQ/SF和SQL-Gen方法生成的合成数据,共21,496行;3) 从评估集中留出的BQ/SF分析领域适应数据,用于提升模型在特定领域的泛化能力,共3,727行。数据集提供了两种格式的文件:通用的`train.jsonl`(采用SWIFT格式)和专门适配FlagEmbedding库训练流程的`train_flagembedding_format.jsonl`(通常包含查询query、正例pos、负例neg等字段)。该数据集主要适用于文本到SQL(text-to-sql)、模式链接(schema-linking)、信息检索等任务,特别是需要学习句子或文本片段之间相似性关系的对比学习场景。
This dataset, named embedding-0.6b-spider2.0-v2-data, is a specialized training dataset for contrastive learning, designed to serve the retriever/reranker models in the GRAST-SQL Stage-I phase and the specific pre-trained embedding model (`thanhdath/embedding-0.6b-spider2.0-v2`). The total dataset size is 42,965 rows, and it explicitly excludes SynSQL data. The data content is a mixture of three main components: 1) training data from the BIRD and Spider benchmark datasets, totaling 17,742 rows; 2) synthetic data generated based on Spider 2.0 using BQ/SF and SQL-Gen methods, totaling 21,496 rows; 3) BQ/SF analysis domain adaptation data held out from the evaluation set to enhance the models generalization ability in specific domains, totaling 3,727 rows. The dataset provides two file formats: the general `train.jsonl` (using SWIFT format) and the `train_flagembedding_format.jsonl` specifically adapted for the FlagEmbedding library training pipeline (typically containing fields such as query, positive example pos, and negative example neg). This dataset is primarily suitable for tasks such as text-to-SQL, schema-linking, and information retrieval, especially in contrastive learning scenarios that require learning similarity relationships between sentences or text fragments.
数据集概述
数据集名称:embedding-0.6b-spider2.0-v2-data
任务类别:特征提取、句子相似度
标签:text-to-sql, schema-linking, retrieval, spider2, bird, contrastive
数据集用途:用于GRAST-SQL第一阶段检索器/重排序器以及模型thanhdath/embedding-0.6b-spider2.0-v2的对比训练数据。
数据集规模:共42,965行,不包含SynSQL。该数据集是标准的Stage-I v2版本,两个文件(train.jsonl和train_flagembedding_format.jsonl)均包含相同的42,965行。
数据集构成:
| 数据组件 | 行数 |
|---|---|
| BIRD训练集 + Spider训练集 | 17,742 |
| Spider 2.0合成数据(BigQuery/Snowflake + SQL-Gen) | 21,496 |
| BigQuery/Snowflake分析领域自适应数据(从评估中保留) | 3,727 |
文件格式:
train.jsonl:SWIFT格式train_flagembedding_format.jsonl:包含query/pos/neg等字段的格式




