BCCard/BCAI-Finance-Kor-Embedding-Pair

1. Overview

Korean financial-domain (anchor, positive) pair dataset for sentence-embedding model training and as a retrieval corpus for evaluation.

This dataset acts in two roles:

1. Training input - 45,589 (anchor, positive) pairs for in-batch / hard-negative contrastive learning.
2. Retrieval corpus - source_chunk_id (≈36k unique chunks after q-variant dedupe) is the candidate pool for IR evaluation in the companion BCCard/BCAI-Finance-Kor-Embedding-Triplet dataset.

1.1. TL;DR

• 45,589 records - 1 row = (anchor, positive) pair with full chunker metadata
• 36,281 unique source_chunk_id - the retrieval corpus (after dedupe of q1/q2/q3 variants)
• Mixed sources - BCAI Q&A (BCCard's internal QA) + BC Card website crawl (cards, FAQ, notice, terms)
• Multi-source chunking - LumberChunker (LLM-guided semantic boundaries) + Multi-Query Instruction Backtranslation (q1/q2/q3 variants)

1.2. Schema

(1) Columns

column	type	description
anchor	string	query text - either the raw chunk text (training) or its instruction-backtranslated variant (q1/q2/q3)
positive	string	doc text (chunker output - the chunk content itself)
source	string	base (BCAI raw) / synth (LLM-processed crawl or BCAI)
chunk_id	string	unique row id (includes q_variant suffix)
source_chunk_id	string	chunk-level source id (q-variant stripped) - this is the corpus entry key
chunk_method	string	base_skip / base_long_llm / llm_pipeline / faq_direct
chunk_index	int	sub-chunk index within source_chunk_id
total_chunks	int	total sub-chunks for this source_chunk_id
query_variant	float	1.0 / 2.0 / 3.0 (synth q-variant) or null (raw)
section_id / context_summary	string	LumberChunker metadata
chunking_config_hash	string	chunker config hash (reproducibility)
page_type / source_url / title / breadcrumb	string	crawl metadata (null for BCAI base)

(2) source_chunk_id - the corpus key

After dedupe of q-variants (1 product = 1 doc text, 3 query variants per chunk), the unique source_chunk_id set is the retrieval corpus:

crawl_..._credit-card_29000   ← LumberChunker sub-chunk 0 of product 29 (= source)
crawl_..._credit-card_29008   ← sub-chunk 8 of product 29
base_bcai_011324_000          ← BCAI raw row 11324 (no chunking)
crawl_..._faq_197000          ← FAQ direct Q&A (no LLM, self-source)

• Trailing number multiple of 1000 = product/article boundary (for crawl_*)
• Trailing number = src_idx * 1000 + sub_chunk_idx (combined encoding by LumberChunker)
• Companion Triplet dataset's relevant_chunk_ids references these source_chunk_id values

1.3. Quick Start

(1) As a corpus for retrieval evaluation

from datasets import load_dataset

pair = load_dataset("BCCard/BCAI-Finance-Kor-Embedding-Pair", split="train")

# Build the corpus dict (key = source_chunk_id, value = doc text)
corpus = {}
for r in pair:
    corpus.setdefault(r["source_chunk_id"], r["positive"])

print(f"Corpus size: {len(corpus):,} unique chunks")        # 36,281

# Use with the companion Triplet dataset for IR evaluation
triplet = load_dataset("BCCard/BCAI-Finance-Kor-Embedding-Triplet")
for row in triplet["test"]:
    qrels = dict(zip(row["relevant_chunk_ids"], row["relevance_scores"]))
    # rank corpus by similarity to row["anchor"] / compute NDCG, MRR, Recall, P@1

(2) As contrastive training data

from datasets import load_dataset
from sentence_transformers import SentenceTransformer, losses
from sentence_transformers.training_args import SentenceTransformerTrainingArguments
from sentence_transformers.trainer import SentenceTransformerTrainer

pair = load_dataset("BCCard/BCAI-Finance-Kor-Embedding-Pair", split="train")

# Use anchor / positive only (in-batch negatives)
model = SentenceTransformer("Qwen/Qwen3-Embedding-0.6B")
loss = losses.MultipleNegativesRankingLoss(model)

# Or use with the companion Triplet dataset (with explicit hard negatives)
triplet = load_dataset("BCCard/BCAI-Finance-Kor-Embedding-Triplet", split="train")
# triplet has (anchor, positive, negative) - explicit hard negatives from FAISS + LLM judge

1.4. Methodology

(1) Data pipeline (hopper module of MokaGrinder)

BCAI Q&A + BC Card site crawl (4 sites: personal/merchant/corporation/VIP)
   ↓ structuring + text-noise removal + PII masking
   ↓ LumberChunker (LLM-guided semantic chunking; dual cap 512/1024 tokens)
   ↓ Multi-Query Instruction Backtranslation (3 query variants per chunk)
   ↓ Cross-source dedupe + integration
This dataset (Pair pool)
   ↓ Hybrid Hard-Negative Mining (FAISS top-K + Claude Sonnet judge)
Companion Triplet dataset (BCCard/BCAI-Finance-Kor-Embedding-Triplet)

(2) Chunking policy

• LumberChunker - LLM (Claude Sonnet) detects semantic boundaries within long pages
• Dual cap - pass-through chunks ≤ 512 tokens, LLM-based chunks ≤ 1024 tokens (aligned with MAX_SEQ_LENGTH=1024)
• faq_direct - FAQ Q&A pairs bypass LLM (already discrete units)
• Multi-Query Instruction Backtranslation - 3 query variants per chunk for instruction-tuned diversity

2. Evaluation

2.1. Distribution Stats

(1) Records per source prefix

chunk_id prefix	records	unique source_chunk_id	description
base_*	31,048	31,048	BCAI raw Q&A rows (LLM ✗, 1 row = 1 chunk)
synth_base_*	1,737	579	BCAI rows processed by LumberChunker + q-variants
synth_crawl_*	12,226	4,076	BC Card crawl processed by LumberChunker + q-variants
crawl_* (FAQ direct)	578	578	BC Card FAQ Q&A direct (LLM ✗, 1 Q&A = 1 source)
total	45,589	36,281

2.2. Quality Self-Assessment (LLM-assisted audit)

The dataset was audited with Claude Fable 5 (2026-06-12): integrity checks (id uniqueness / empty rows / degenerate rows / q-variant consistency / metadata sanity), corpus duplicate-document detection, and cross-dataset referential integrity against the companion Triplet dataset.

Overall score: 82 / 100 - scored against the bar for public datasets, on the same rubric as the companion Triplet dataset (78/100). Pair scores higher because it carries no relevance labels - the Triplet's weakest pillar does not apply here.

dimension	score	rationale
Data integrity	9.5/10	all 45,589 chunk_id unique; 0 empty values; 0 degenerate rows (anchor==positive); chunk_index / total_chunks consistent
Corpus quality	8/10	36,281 unique source_chunk_id, exactly 1 doc text per id (q-variant consistent: 0 mismatches); deducted for 1.06% exact-duplicate doc texts (finding below)
Cross-dataset referential integrity	10/10	every relevant_chunk_ids reference in the companion Triplet qrels (validation 830 / test 816 unique chunks) resolves in this corpus - 0 missing
Composition & metadata	9/10	source distribution matches this card (base 32,785 / synth 12,804); full chunker metadata incl. chunking_config_hash for reproducibility
Documentation	9/10	Limitations with measured numbers + Future Work

(1) Audit finding - corpus exact-duplicate documents

• 1.06% of the corpus (386 chunks in 189 groups) share byte-identical doc text under different source_chunk_id - the same content published on multiple pages (same root cause as the cross-split duplicate queries reported in the companion Triplet dataset)
• Impact on IR evaluation: qrels label specific source_chunk_id values, so a model that retrieves an unlabeled twin of a relevant chunk is scored rel=0 despite returning the right content (unjudged-relevant)
• Remedy is shared with the Triplet's pooling-based qrels enrichment plan (duplicate twins get judged and labeled); a corpus-side dedupe by doc text is the alternative for a future version

2.3. Limitations

• Domain: Korean financial services (BC Card products, FAQs, terms, BCAI Q&A) - narrow domain; generalization to other domains not guaranteed
• Query style heterogeneity: base anchors are real user questions while synth anchors are LLM-generated queries - the two have measurable stylistic differences (see companion Triplet dataset for surface-signal measurements)
• Corpus duplicate documents: 1.06% of corpus chunks are byte-identical twins under different ids (see audit finding above) - relevant for IR evaluation against this corpus

3. Future Work

3.1. Shared upgrades with the companion Triplet dataset

Planned upgrades shared with the companion Triplet dataset (surface-signal suppression + semantic-signal enrichment):

1. Query style diversification - reduce the stylistic gap between base (real user questions) and synth (LLM-generated queries) via paraphrase / backtranslation
2. Lexically-hard negatives (Triplet v4) - BM25-mined negatives built on top of this corpus to close the lexical-overlap shortcut
3. LLM-judge label verification - sample-audit of the distance-based graded relevance labels used by the companion Triplet dataset

4. Meta Info

4.1. Citation

@dataset{bccard_bcai_finance_kor_embedding_pair,
  title  = {BCAI Finance Kor Embedding Pair},
  author = {BC Card AX Team},
  year   = {2026},
  url    = {https://huggingface.co/datasets/BCCard/BCAI-Finance-Kor-Embedding-Pair},
  note   = {Companion evaluation dataset: BCCard/BCAI-Finance-Kor-Embedding-Triplet}
}

4.2. See Also

• Triplet (queries + graded labels): BCCard/BCAI-Finance-Kor-Embedding-Triplet
• Pipeline code: MokaGrinder (MoG)