BCAI Finance Kor Embedding Triplet: Korean Financial Text Triplets

45,394 triplets of Korean financial text for fine-tuning sentence-embedding models, with graded relevance labels. The dataset was created by BCCard/BCAI using FAISS top-K and Claude Sonnet LLM judge for hybrid hard-negative mining. It was last updated on June 11, 2026.

Use Cases

• Fine-tuning sentence-embedding models based on the described triplet structure (anchor/positive/negative).
• Evaluating model performance using the graded relevance labels provided for validation and test splits.
• Training retrieval models for financial document matching, as it is a companion to a retrieval corpus dataset.

Strengths

• 45,394 total triplets provide a substantial training corpus.
• The dataset includes graded relevance labels for validation and test splits, enabling nuanced evaluation.
• Triplets were mined using a hybrid method involving FAISS and a large language model judge.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for individual splits beyond the total is unknown, which may limit suitability assessment.
• Data may reflect bias inherent to its source and construction method on the Hugging Face platform.

Provenance

Source

BCCard

Collection Method

Mined with FAISS top-K and Claude Sonnet LLM judge (Hybrid Hard-Negative mining).

Freshness

Last updated 2026-06-11 06:52:29.