BCCard/BCAI-Finance-Kor-Embedding-Triplet
收藏资源简介:
BCAI金融韩语嵌入三元组数据集是一个专门用于微调句子嵌入模型的韩语金融领域数据集。它包含45,394个三元组(锚点、正例、负例),通过FAISS top-K和Claude Sonnet LLM判断进行混合硬负例挖掘生成。数据集分为训练集(43,394个)、验证集(1,000个)和测试集(1,000个,作为黄金集)。在anchor_parent_doc_id级别上进行了源分离,以防止跨分割的兄弟泄漏。验证集和测试集包含分级相关性标签(relevant_chunk_ids和relevance_scores),相关性分数分为1、2、3级(0表示稀疏)。数据来源于BCAI问答和BC Card网站爬取,经过结构化、文本噪声处理、PII掩码、LLM引导的语义分块、多查询指令反向翻译等处理。该数据集主要用于检索评估,支持计算NDCG@10、MRR@10、Recall@10和P@1等指标。数据集还提供了详细的局限性分析(如领域狭窄、硬负例泄漏、词汇重叠偏差等)和未来工作计划(如词汇硬负例挖掘、查询风格多样化等)。
The BCAI Finance Kor Embedding Triplet dataset is a Korean financial-domain triplet dataset designed for fine-tuning sentence-embedding models. It contains 45,394 triplets (anchor/positive/negative) mined using FAISS top-K and Claude Sonnet LLM judge (Hybrid Hard-Negative mining). The dataset is split into three parts: train (43,394), validation (1,000), and test (1,000, the Golden set). It is source-disjoint at the anchor_parent_doc_id level to prevent sibling leakage across splits. The validation and test splits include graded relevance labels (relevant_chunk_ids and relevance_scores) with relevance scores of 1, 2, 3 (0 for sparse). The data is sourced from BCAI Q&A and BC Card site crawls, processed through structuring, text-noise handling, PII masking, LLM-guided semantic chunking, and multi-query instruction backtranslation. The dataset is primarily used for retrieval evaluation, with recommended metrics including NDCG@10, MRR@10, Recall@10, and P@1. It also includes detailed limitations (e.g., narrow domain, hard-negative leakage, lexical overlap bias) and future work plans (e.g., lexically-hard negatives, query style diversification).




