--- license: cc-by-nc-4.0 language: - en tags: - synthetic - finance - trading - knowledge-base - instruction-tuning - rag - reranking - adversarial-robustness pretty_name: "Trading Concepts Knowledge Base (FIN-QA-001) — Sample" size_categories: - 1K **Positioning note.** Answers are *templated, distributionally-controlled prose* > rendered from a concept ontology — not human-verified factual ground truth. > This corpus is built for **structural / retrieval / reranker / adversarial-robustness** > work (topic & difficulty calibration, depth-tier ranking, refusal behavior), > not for teaching factual financial knowledge via supervised fine-tuning. See > **Limitations** below. Each item is **educational** and **not investment advice**. ## What's a "depth tier" here (not a gold/distractor scheme) Unlike a gold/silver/distractor design, every question in `multi_depth_answers` has **exactly four answers**, one per audience depth: `beginner`, `intermediate`, `advanced`, `institutional`. All four are legitimate renderings of the same concept at different sophistication levels. There is intentionally **no single "correct" answer** — the structure supports depth-conditioned generation and depth-ranking tasks. ## Calibration anchors | Metric | Observed (seed 42) | Target range | Anchor | |---|---|---|---| | Scenario question share | 0.246 | 0.18–0.30 | Design topic mix (scenario-weighted) | | Difficulty mean (1–5) | 3.07 | 2.70–3.30 | Bloom's Taxonomy difficulty centering | | Max persona share | 0.170 | 0.10–0.22 | CFA / advisor persona breadth | | Max L1 topic share | 0.128 | 0.08–0.20 | CFA curriculum topic breadth (12 domains) | | Adversarial refusal/clarify fraction | 0.467 | 0.34–0.56 | OWASP LLM adversarial behavior target | | Prompt-injection share | 0.250 | 0.12–0.34 | OWASP LLM01 | | Reasoning-required answer share | 0.682 | 0.60–0.80 | FinQA reasoning-required fraction | | Formula-coverage share | 0.068 | 0.02–0.18 | Quantitative-formula coverage | Heavily-weighted **structural integrity floors** (all exact, all pass): exactly 4 depth answers per question; all 4 depth tiers present & distinct; full referential integrity across FKs; exactly one misconception per question; per-table column-count contract; no relation self-loops. ## Tables (schema highlights) | Table | Rows (sample) | Key columns | |---|---|---| | `finqa001a_qa_pairs` | 500 | `qa_id`, `concept_id`, `question_text`, `question_type`, `persona_type`, `difficulty` | | `finqa001a_multi_depth_answers` | 2,000 | `answer_id`, `qa_id`, `depth_level`, `answer_text`, `contains_formula_flag`, `requires_reasoning_flag` | | `finqa001a_misconceptions` | 500 | `misconception_id`, `concept_id`, `incorrect_statement`, `why_wrong`, `correct_explanation`, `error_type` | | `finqa001a_adversarial_queries` | 120 | `adv_id`, `qa_id`, `attack_type`, `adversarial_question`, `expected_behavior` | | `finqa001a_conversations` | 80 | `conv_id`, `persona_type`, `turn_sequence` (JSON), `topic_drift_flag`, `resolution_flag` | | `finqa001a_concepts` | 250 | `concept_id`, `concept_name`, `category_l1`, `category_l2`, `difficulty_level`, `institutional_relevance_score`, `description_short`, `description_long` | | `finqa001a_ontology` | 434 | `node_id`, `node_type`, `name`, `parent_node`, `depth_level` | | `finqa001a_relations` | 500 | `relation_id`, `source_node`, `target_node`, `relation_type`, `strength_score`, `relation_description` | | `finqa001a_formulas` | 17 | `formula_id`, `concept_id`, `formula_latex`, `variable_definitions`, `interpretation` | | `finqa001a_examples` | 250 | `example_id`, `concept_id`, `example_type`, `example_description`, `solution_steps` | | `finqa001a_use_cases` | 250 | `use_case_id`, `concept_id`, `use_case_type`, `use_case_description` | | `finqa001a_personas` | 10 | `persona_id`, `persona_name`, `risk_tolerance`, `knowledge_level`, `typical_query_style`, `institutional_flag` | `source_node` / `target_node` in `relations` reference `concept_id` values. `conversations.turn_sequence` is a JSON-encoded list of turn objects. ## Loading ```python import pandas as pd qa = pd.read_csv("finqa001a_qa_pairs.csv") answers = pd.read_csv("finqa001a_multi_depth_answers.csv") # Join: each question -> its 4 depth-tiered answers merged = qa.merge(answers, on="qa_id") print(merged.groupby("qa_id").size().value_counts()) # all == 4 ``` ```python from datasets import load_dataset qa = load_dataset("xpertsystems/fin-qa-001-sample", "qa_pairs")["train"] answers = load_dataset("xpertsystems/fin-qa-001-sample", "multi_depth_answers")["train"] ``` ## Use cases - **Supervised fine-tuning (style/format/depth):** train depth-conditioned answer generation (beginner vs institutional voice) over a fixed concept set. - **Preference / ranking data:** build depth-preference pairs (e.g. prefer institutional over beginner for an expert persona) for reranker / RLHF-style signals — note pairs encode *audience fit*, not factual correctness. - **RAG & reranker evaluation:** topic-calibrated `(query, answer-shape)` pairs with controlled difficulty and 12-domain spread for MRR/NDCG-style metrics. - **Adversarial robustness:** `adversarial_queries` provides prompt-injection, misleading, oversimplification, and hallucination-bait probes with `expected_behavior` labels for refusal/clarify evaluation. ## Limitations (full disclosure) The build process inspected the engine line-by-line. Disclosed observations: 1. **Answers are templated prose, not verified facts.** Answer text is rendered from concept metadata and persona/depth templates. It is plausible and structurally complete, but is **not** human-verified financial truth. Do not use `(question, answer)` pairs as factual SFT ground truth. 2. **No train/test split column is emitted.** The engine defines a `TRAIN_CUTOFF_ID = 16000` constant (exported in the manifest), but at sample scale (max `qa_id` = 500) **all rows fall below the cutoff** and no split column is written to any table. Consumers must construct their own splits. 3. **No gold/preferred tier.** This is a **depth-tier** corpus (4 audience levels), intentionally without a single "correct" answer. If your pipeline expects a gold/distractor schema, this is a different structure. 4. **Misconception count is gated by concept count.** The engine writes misconceptions only while `qa_index <= n_concepts * 2`. The sample is built at `n_concepts = 250`, so all 500 questions receive exactly one misconception. In the full 20,000-question product (`n_concepts = 1000`) only the first ~2,000 questions carry a misconception. 5. **`QUESTION_TYPE_WEIGHTS` is not explicitly renormalized** in engine code; it already sums to 1.0, so behavior is correct, but it is not defensively normalized like the other weight vectors. 6. **Single shared RNG, order-dependent.** All randomness draws from one `np.random.default_rng(seed)` consumed sequentially. This is fully deterministic per seed but means table contents are coupled to generation order; there are no per-record sub-RNGs. 7. **`manifest.json` contains a wall-clock `generated_at` timestamp**, which is excluded from determinism verification (data files are byte-identical per seed; the timestamp is not). No benchmark-theater was found: no hardcoded validation values, no `max(actual, target)` floors, no always-true passes, no referential-integrity leaks. Scorecard ranges were calibrated to observed 6-seed behavior; the heavy weight sits on exact structural floors. ## Sample vs. full product | Dimension | Sample (this repo) | Full product | |---|---|---| | Questions | 500 | 20,000 | | Concepts | 250 | 1,000 | | Multi-depth answers | 2,000 | 80,000 | | Misconceptions | 500 | ~2,000 | | Adversarial queries | 120 | ~3,000 | | Conversations | 80 | 3,000 | | License | CC-BY-NC-4.0 | Commercial | | Validation | 6/6 seeds Grade A+ (10.0/10) | Full-scale QA suite | ## Determinism Re-running the engine with the same seed produces **byte-identical data files** (verified across all 12 CSVs) and **identical scored metrics**. The only per-run variation is the manifest `generated_at` wall-clock timestamp, which is excluded from the determinism check. ## Citation ```bibtex @misc{xpertsystems_finqa001_2026, title = {Trading Concepts Knowledge Base (FIN-QA-001): A Synthetic, Ontology-Driven Multi-Depth Trading Q&A Corpus}, author = {XpertSystems.ai}, year = {2026}, howpublished = {Hugging Face Datasets}, note = {Sample (500 questions) of a 20,000-question commercial product. Difficulty mix calibrated to Bloom's Taxonomy; topic breadth to the CFA Institute curriculum; reasoning-required fraction to FinQA-style financial-reasoning corpora; adversarial attack-type and refusal mix to the OWASP LLM Top-10. License CC-BY-NC-4.0.}, url = {https://xpertsystems.ai} } ``` Anchored benchmarks referenced for calibration: Bloom's Taxonomy of educational objectives; CFA Institute candidate body of knowledge (topic-area weights); FinQA (Chen et al., financial numerical-reasoning QA); OWASP Top-10 for LLM Applications (adversarial attack taxonomy). ## Disclaimer This dataset is **synthetic** and provided for AI/ML research and engineering. Its content is **educational and illustrative only** and is **not investment advice**, not a recommendation to buy or sell any security, and not a substitute for professional financial, legal, or compliance guidance.