ardge-hf/crag-finance-rag

CRAG Finance RAG

Finance-domain QA pairs and a matching knowledge base, derived from CRAG (Comprehensive RAG Benchmark) Task #3. Built for testing/benchmarking a RAG pipeline end-to-end: retrieve from knowledge, answer against QA.

Two splits, different schemas — read whichever split you need directly, they are not meant to be concatenated.

QA split (661 rows)

The finance-domain subset of CRAG's QA pairs (validation + test merged into one split).

column	type	description
interaction_id	string	unique id
query_time	string	when the question was asked
domain	string	always finance here
question_type	string	simple / comparison / aggregation / set / false_premise / multi-hop / post-processing / simple_w_condition
static_or_dynamic	string	static / slow-changing / fast-changing / real-time
query	string	the question
answer	string	ground-truth answer
alt_ans	list[string]	acceptable alternative answers
popularity	string	head/torso/tail = answer comes from the KG; empty = answer comes from the web (per CRAG's own field definition)

knowledge split (28,366 rows)

A finance knowledge base built from two sources, meant to be embedded/indexed for retrieval:

column	type	description
company_description	string	markdown document
embedding	list[float32], dim 1024	Qwen/Qwen3-Embedding-0.6B embedding of the document

It mixes two kinds of documents, not distinguished by any column — the title line of each markdown document is the only signal:

• 7,315 rows — one company per row. Rebuilt from CRAG's finance mock KG (company_name.dict + 5 sqlite tables: price/detailed-price/dividend history, market cap, financial info), keyed by ticker symbol. Each row: company overview, key financial metrics, dividend history, price history (summary + full daily table), intraday price summary. 405 tickers have no financial data beyond name+symbol (very short rows, ~22 chars).
• 21,051 rows — one web page per row. Every unique web page from CRAG Task #3's search_results for finance-domain questions, converted from raw HTML to markdown with trafilatura. Included regardless of relevance to any company or question — most of it is generic finance content (retirement-account comparisons, historical events, definitions, market-wide lists) that doesn't answer any specific question in the QA split. This is intentional noise: it exists to stress-test retrieval precision, not to be a curated corpus. 1,399 pages failed extraction (empty/unparseable) and were dropped.

Notes

• Embedding truncation: Qwen/Qwen3-Embedding-0.6B has a 32,768-token context. 71 documents exceeded it (up to 1.2M characters — e.g. full SEC filings pulled in verbatim by a search result). For those, embedding was computed on a truncated prefix (shrunk from 50K down to 1K characters until it fit); company_description itself is not truncated, so retrieval on the full text is unaffected — only the embedding of the tail may be less representative for a handful of documents.
• Company-to-webpage linking was deliberately not attempted for this split. An earlier iteration matched ticker symbols found in page URLs/titles (against a blocklist of false positives like /terms/v/... glossary slugs) to fold matched pages into their company's row — see this dataset's commit history if you want that version's logic. This release keeps the two document types unmerged so relevance judgment is left entirely to the retriever being tested.
• All embeddings were produced by direct inference against a live vLLM endpoint (Qwen/Qwen3-Embedding-0.6B, 1024-dim) — not precomputed by a third party.

License

CC BY-NC 4.0, inherited from the source CRAG dataset (non-commercial use).

Citation

@article{yang2024crag,
  title={CRAG--Comprehensive RAG Benchmark},
  author={Yang, Xiao and Sun, Kai and Xin, Hao and Sun, Yushi and Bhalla, Nikita and Chen, Xiangsen and Choudhary, Sajal and Gui, Rongze Daniel and Jiang, Ziran Will and Jiang, Ziyu and others},
  journal={arXiv preprint arXiv:2406.04744},
  year={2024}
}