xpertsystems/fqa-000-sample

FQA-001 — Financial Q&A Dataset: Investment-Domain Questions with Tiered Answers, Conversations & Adversarial Variants (Sample)

Synthetic investment-domain question-and-answer corpus with controlled difficulty, topic distribution, tiered answers, multi-turn conversations, and adversarial variants. Questions span 12 CFA curriculum domains (Equities, Fixed Income, Derivatives, Portfolio Management, Economics, Alternative Investments, Corporate Finance, Risk Management, Regulation & Ethics, Quantitative Methods, Real Assets, Behavioral Finance) across 8 structural formats (definition, calculation, comparison, scenario, what-if, error-spotting, best-of-N, regulatory-applicability).

This is a 500-question sample of the full FQA-001 product (50,000 questions). It is synthetic — generated by a deterministic, benchmark-anchored engine. It contains no real financial advice, securities data, or PII. Tickers, entities, and prices are fictional.

Not investment advice. This dataset is for ML training/evaluation, instruction-tuning, retrieval/QA benchmarking, and research only. It must not be used as a source of financial, investment, legal, or regulatory advice.

Unit of observation

The unit is the question. Each question has 4 answers (1 gold + 1 silver + 2 plausible-wrong distractors). Tables key on question_id / answer_id / entity_id. The sample is entirely within the engine's train split (chronological cutoff at question_id 40000); the train_test_split column is retained for schema fidelity.

Calibration anchors

Sample-level observed values (seed 42, 500 questions):

Metric	Observed	Target	Anchor
CFA L1 topic abs deviation	0.12	≤0.22	CFA Level-I topic-area weights (12 domains)
Bloom mean level	2.90	2.7–3.2	Bloom's taxonomy CFA skew (~2.95)
Numerical-question fraction	~0.40	0.20–0.55	calculation/scenario template share
Adversarial attack-mix abs deviation	0.16	≤0.30	TruthfulQA + AdvGLUE + OWASP LLM Top 10
Adversarial refusal fraction	~0.59	0.40–0.65	pii_probe + jailbreak + prompt_injection
Conversation resolution rate	~0.72	0.60–0.84	Accenture financial-chatbot benchmark
Entity top-10% reference share	high	≥0.18 (floor)	Zipf(s=1.2) entity reuse (Reuters corpus)
Gold-per-question violations	0	=0 (floor)	exactly one gold answer
Answers-per-question violations	0	=0 (floor)	exactly four answers
Gold error-type violations	0	=0 (floor)	gold carries error_type='none'
QA-pair integrity violations	0	=0 (floor)	valid gold id + 3 distractors

Validation: Grade A+ (10.00/10) across all six canonical seeds (42, 7, 123, 2024, 99, 1). All eight data CSVs are byte-for-byte deterministic per seed.

Tables

Eight relational CSVs + manifest:

• fqa_001_questions.csv — 500 questions × 22 cols: text, topic L1/L2, Bloom level, difficulty, format type, persona, entity refs, numerical/multi-hop/temporal/negation flags, token/sentence counts, train/test split.
• fqa_001_answers.csv — 2,000 answers × 8 cols: answer text, quality tier (gold/silver/plausible_wrong), is_gold, error type, error severity, token count.
• fqa_001_qa_pairs.csv — 500 rows × 9 cols: gold answer id, distractor id JSON array, context snippet, topic, Bloom, difficulty.
• fqa_001_entities.csv — 500 entities × 10 cols: synthetic ticker/name, sector, asset class, OU price anchor, credit rating.
• fqa_001_topic_taxonomy.csv — 129 rows × 7 cols: L1/L2 hierarchy with CFA weights and query-frequency indices.
• fqa_001_personas.csv — 10 personas × 7 cols: knowledge level, risk tolerance, query style.
• fqa_001_conversations.csv — 200 conversations × 8 cols: multi-turn dialogues (Markov topic drift), resolution flag, turn sequence JSON.
• fqa_001_adversarial_pairs.csv — 200 rows × 11 cols: attack type, adversarial text, expected-refusal flag, jailbreak/injection/hallucination scores.
• fqa_001_manifest.json — metadata, row counts, seed, parameters, split definition.

Loading

import pandas as pd, json
questions = pd.read_csv("fqa_001_questions.csv")
answers   = pd.read_csv("fqa_001_answers.csv")
qa_pairs  = pd.read_csv("fqa_001_qa_pairs.csv")

# Assemble a gold-answer training pair
row = qa_pairs.iloc[0]
gold = answers.loc[answers.answer_id == row.gold_answer_id, "answer_text"].iloc[0]
q    = questions.loc[questions.question_id == row.question_id, "question_text"].iloc[0]
print(q, "->", gold)

from datasets import load_dataset
questions = load_dataset("xpertsystems/fqa-001-sample", "questions")

Use cases

• Instruction-tuning / SFT on tiered (gold/silver/distractor) financial QA.
• Reward-model / preference data (gold > silver > plausible_wrong).
• Retrieval-QA and multiple-choice (best-of-N) benchmarking.
• LLM safety / adversarial-robustness evaluation (prompt injection, jailbreak, PII probe, hallucination bait, out-of-scope).
• Multi-turn dialogue and topic-drift modeling.
• Difficulty-calibrated (Bloom) curriculum learning and topic-coverage analysis.

Limitations (honestly disclosed)

• Synthetic, template-generated text; not investment advice. Questions and answers are produced from deterministic template pools with sampled numerics and fictional entities. Answer prose is structurally coherent but not authored or fact-checked by finance professionals; gold answers are "correct by construction" within the template logic, not verified against real markets.
• Taxonomy count note. The engine's header comment describes "200" L2 subtopics, but the actual taxonomy defines 129 subtopics across 12 L1 domains. The manifest row count is computed correctly from the taxonomy (129), so the data is internally consistent; only the header comment overstates the figure.
• Sample is train-split only. The engine's chronological train/test split cuts at question_id 40000, so a 500-question sample is entirely "train". The column is retained for schema fidelity; the full product spans both splits.
• Adversarial mix is seed-stable. Per-record sub-RNGs (answers, conversations, adversarial pairs) are seeded from their record id, so the adversarial attack-type distribution is consistent across seeds (well-calibrated to target weights). Question and entity content still varies by seed. This is a determinism design choice, not a defect.
• Marginal calibration, not full joint fidelity. Topic/Bloom/adversarial/error-tier distributions and the structural integrity guarantees are anchored; higher-order correlations (e.g. topic × difficulty × persona joint structure) beyond what the templates encode are not independently validated.
• Small-sample variance. At 500 questions the 12-domain topic and 5-category adversarial distributions carry multinomial sampling noise; scorecard bands accommodate this and the structural integrity floors are weighted to dominate.

Commercial / full version

	Sample (this)	Full (commercial)
Questions	500	50,000
Answers	2,000	200,000
Entities	500	5,000
Conversations	200	5,000
Adversarial pairs	200	5,000
Topic taxonomy	129 subtopics	129 subtopics
Train/test split	train only	train (≤39999) + test (≥40000)
Bloom skew / topic regime	CFA / balanced	cfa|uniform|advanced × balanced|equity_heavy|macro_heavy
Formats	CSV + manifest	CSV / JSON + manifest
Seeds / reproducibility	6 canonical	Unlimited
License	CC-BY-NC-4.0	Commercial
Support	—	SLA, custom topic regimes, adversarial intensity tuning

Contact pradeep@xpertsystems.ai · https://xpertsystems.ai

Citation

@dataset{xpertsystems_fqa001_2026,
  title  = {FQA-001: Synthetic Financial Q&A Dataset --- Investment-Domain Questions
            with Tiered Answers, Multi-Turn Conversations & Adversarial Variants (Sample)},
  author = {XpertSystems.ai},
  year   = {2026},
  publisher = {Hugging Face},
  note   = {Synthetic data. Not investment advice. Fictional entities. Calibration
            anchors: CFA Institute Level-I curriculum topic-area weights; Bloom's
            taxonomy item-difficulty skew; FinQA (Chen et al. 2021) error analysis;
            TruthfulQA + AdvGLUE + OWASP LLM Top 10 adversarial taxonomy; Reuters
            corpus Zipf entity reuse; Accenture financial-chatbot resolution benchmark.},
  url    = {https://huggingface.co/datasets/xpertsystems/fqa-001-sample}
}