FinQA: A Dataset of Numerical Reasoning over Financial Data

“For all models, we use the Adam optimizer (Kingma and Ba, 2015).”

“Retriever The full page of the financial report can go beyond 2,000 tokens, which cannot be coped with the current popular QA models (Devlin et al., 2019).”

“, 2019) and RoBERTa (Liu et al., 2019), our proposed approach outperforms all other baselines and achieves an execution accuracy of 65.”

“3703 scales linearly with sequence length, like the Longformer (Beltagy et al., 2020).”

“Financial QA is more challenging than classic QA (Rajpurkar et al., 2018; Yang et al., 2018) because it requires the system to spot relevant information across heterogeneous sources, such as tables and unstructured texts, and then create a numerical reasoning path to connect all the information.”

“Financial QA is more challenging than classic QA (Rajpurkar et al., 2018; Yang et al., 2018) because it requires the system to spot relevant information across heterogeneous sources, such as tables and unstructured texts, and then create a numerical reasoning path to connect all the information.”

“This way the model is able to learn the program structures as patterns from very large-scale data (~40k for MathQA), but may fail on learning the reasoning paths.”

“The performance of using FinBert is no better than BERT-large, mostly because its pre-training corpus is limited (~30M words from news articles).”

“Another kind of QA datasets related to ours is the math word problem datasets, like MaWPS (Koncel-Kedziorski et al., 2016), MathQA (Amini et al.”

“Among these, question answering on tabular data (TableQA) is a challenging task that requires the understanding of table semantics, as well as the ability to reason and infer over relevant table cells (Herzig et al., 2021; Chen et al., 2020b, 2021b).”

“Similar to DROP, F IN QA (Chen et al., 2021) consists of numerical reasoning questions over financial data.”

“, FinQA [6] and TAT-QA [29] to test its efectiveness.”

“…Link TabRecSet [188] Wild [191] Wikipedia 2017 24.2k tables, 80k queries Link HybridQA [22] Wikipedia 2020 13k tables, 70k questions Link FinQA [26] Financial reports 2021 2.7k report pages, 8.2k questions Link TAT-QA [232] Financial reports 2021 2.7k hybrid contexts, 16.5k questions Link…”

“Financial LLM benchmarks have evolved from document-understanding frameworks such as FinanceBench (Islam et al., 2023), PIXI-U/FLARE (Xie et al., 2023), FinBen (Xie et al., 2024), FinQA (Chen et al., 2021), ConvFinQA (Chen et al., 2022), TAT-QA (Zhu et al., 2021), and the Chinese-language CFLUE (Zhu et al., 2024) and…”

“TableQA FinQA (Chen et al., 2021b) focuses on numerical reasoning over financial data, containing questions derived from S&P 500 companies’ earnings reports that involve both structured tables and unstructured text, with annotations created by finance experts.”

“(1) Most benchmarks oversimplify SEC filings as flat, isolated text spans, focusing on information extraction, numerical computation over tabular data, and QA restricted to a single section and reporting year [6, 16, 24, 27].”

“This process culminates in a high-quality numerical reasoning dataset, C AL T AB 151 , composed of 151 table samples drawn from TableBench(Wu et al., 2024) (84), FinQA(Chen et al., 2021) (27), TAT-QA(Zhu et al., 2021) (32), and AitQa(Katsis et al., 2022) (8).”

“Recent works such as TechQA (Castelli et al., 2019), FinTextQA (Chen et al., 2024), FINQA (Chen et al., 2022), PubMedQA (Jin et al., 2019), WikiHowQA (Bolotova-Baranova et al., 2023), and AmazonQA (Gupta et al., 2019) underscore the trend toward benchmarks that capture expert curation, automatic…”

“SQL, as a formal language, provides a vehicle for structured reasoning over relational tables; thus, models generating SQL should ideally also support broader forms of table-based question answering (e.g., TabFact [6], WikiTQ [24], FinQA [7]).”

“TAT-QA (Zhu et al., 2021) and FinQA (Chen et al., 2021) builded the numerical reasoning hybrid dataset which comes from the financial field.”

“For SQL queries and arithmetic expressions, we fine-tune a BART [29] model ourselves on SQUALL [45] and Finqa [6], respectively.”

“Chen et al. (2021) and Chen et al. (2022) have introduced FinQA and ConvFinQA, which are complex question-answering and conversational numerical reasoning tasks for financial reports, respectively.”

“Zhu et al. (2021) and Chen et al. (2021) propose hybrid QA datasets, also focusing on answering questions over financial data.”

“…question answering (Chen et al., 2024a; Lu et al., 2023; Zhu et al., 2024) and complex reasoning (Xie et al., 2024a; Lu et al., 2025; Chen et al., 2021; Yoash et al., 2025; Xie et al., 2025; Jiang et al., 2025), thereby facilitating the evaluation of market prediction, numerical…”

“In the past few years, there has also been significant progress in benchmarks for financial reasoning (Chen et al., 2021; Islam et al., 2023; Bigeard et al., 2025; Xie et al., 2024a; Liu et al., 2025a; Choi et al., 2025; Wang et al., 2025b; Li et al., 2025a; Liu et al., 2025b; Hu et al., 2025;…”

“(Maia et al., 2018) For more demanding and realistic scenarios, datasets like FinQA are critical for testing multi-step numerical reasoning (Chen et al., 2021), while TAT-QA evaluates the essential ability to synthesize information from both text and tables (Zhu et al., 2021).”

“While researchers have gradually introduced specialized benchmarks for the financial domain, such as FinEval (Guo et al. 2024), FinQA (Chen et al. 2021), CFLUE (Zhu et al. 2024), CFinBench (Nie et al. 2024) and MME-Finance (Gan et al. 2024), these benchmarks remain predominantly text-centric or image-centric.”

“Compared to FinQA, we focus on KBQA tasks in general rather than the financial domain.”

“Similarly, Chen et al. (2021) proposed FinQA, another QA benchmark on financial hybrid data emphasizing numeracy skills.”

“Tabular QA and Semantic Parsing (Zhu et al., 2021; Chen et al., 2021; Zhong et al., 2017; Pasu-pat and Liang, 2015) concerns question answering in the domain of tabular data, which often involves a large amount of numbers and requires arithmetic aggregations to arrive at the final answer.”

“Examples include financial tables (Zhang et al., 2024a; Chen et al., 2021), healthcare records (Guo et al., 2024; Shi et al., 2024), and scientific data (Smock et al., 2022), all of which are prevalent in critical real-world applications.”

“Domain-specific efforts have also emerged, including PubMedQA [Jin et al., 2019] for biomedical literature and FinQA [Chen et al., 2021] for financial reasoning.”

“Financial question answering benchmark,s including FinQA [2] and ConvFinQA [3], emphasize retrieval and arithmetic computation but do not assess whether models can enforce accounting principles or discriminate among competing rule violations.”

“AI systems based on large language models (LLMs) have demonstrated remarkable proficiency at discrete, well-defined tasks across diverse domains, such as legal reasoning [34, 37], software engineering [41], drug discovery [39, 81], and finance [21, 91].”

“(iii) Document-level Reasoning , where we use the DocMath-Eval benchmark [91] with several financial reasoning datasets such as TAT-QA [94], FinQA [7], MultiHiertt [90], and TAT-HQA [35].”

“Domain-specific benchmarks now better reflect real-world scenarios in fields such as airlines (Katsis et al., 2022) and finance (Zhu et al., 2021; Chen et al., 2022).”

“However, financial tasks such as financial sentiment analysis [2], [3], financial statement summarization [4], [5], and financial question answering [6], [7] still face significant challenges due to the scarcity of available data and the need for extensive domain-specific knowledge [8].”

“Efforts include WebQA (Chang et al., 2022) and MuMuQA (Reddy et al., 2022)—for two-hop cross-modal reasoning with text and images—as well as HybridQA (Chen et al., 2020), OTT-QA (Chen et al., 2021a), FinQA (Chen et al., 2021b), TAT-QA (Zhu et al., 2021), and TANQ (Akhtar et al., 2024), which include text and tables.”

“The datasets utilized for this type of task comprise TabMWP (Lu et al., 2022b), TAT-QA (Zhu et al., 2021), HiTab (Cheng et al., 2021), MultiHiertt (Zhao et al., 2022), FinQA (Chen et al., 2021b), WikiSQL (Zhong et al., 2017), SQA (Iyyer et al., 2017), and WTQ (Pasupat & Liang, 2015).”

“…is applied in financial question answering and solving mathematical word problems, and numerous datasets have been developed for these purposes (Amini et al., 2019; Cobbe et al., 2021; Patel et al., 2021; Zhu et al., 2021; Chen et al., 2022; Liu et al., 2023; Zhao et al., 2023b; Lu et al., 2023).”

“Later benchmarks further require free-form answer generation (Nan et al., 2022), multi-hop reasoning (Chen et al., 2021a; 2020) and mathematical reasoning (Zhu et al., 2021; Chen et al., 2021b; Lu et al., 2023).”

“…SST-5 for 2 and 5-way sentiment classification respectively (Socher et al., 2013), MNLI for natural language inference (Williams et al., 2017), TREC for 6-way question classification (Li and Roth, 2002), MedQA (Jin et al., 2021) and FinQA for domain-specific question-answering (Chen et al., 2021).”

“Specifically, we collect templates of arithmetic expressions from the Finqa dataset proposed in [28].”

“Early work focused on classification and extraction (Araci, 2019), while recent datasets target numerical reasoning over tables (FinQA (Chen et al., 2021), TAT-QA (Zhu et al., 2021)), multi-turn dialogue (ConvFinQA (Chen et al., 2022)), and chain-of-thought verification (FinChain (Xie et al.,…”

“Similarly on the FinQA benchmark, LLaMA-7B reduces MI avg from 0 .”

“…standardized prediction to trading benchmarks that couple reasoning tasks such as FinQA, FinanceBench, and AlphaFin with open and time safe market simulators and with safety audits that detect risky tool use (Chen, Chen, et al., 2021; Islam et al., 2023; Li, Li, et al., 2024; Yuan et al., 2024).”

“FinQA [148] introduced a series of expert-written QA pairs on financial reports, focusing on numerical reasoning with both textual and tabular evidence.”

“Input modality Expert level reasoning Language FinQA (Chen et al., 2021) Text, Table En ConvFinQA (Chen et al., 2022b) Text, Table En FinanceBench (Islam et al., 2023) Text En FinanceMATH (Zhao et al., 2024) Text, Table En FAMMA (Xue et al., 2025) Text, Image En AuditBench (Wang et al., 2024) Text,…”

“TableQA We evaluate our proposed CoRE on two table-based reasoning datasets, including WikiTable-Question (Pasupat and Liang, 2015a), and FinQA (Chen et al., 2021a).”

“…(1) Limited structural coverage: Text-only datasets such as TaT-QA (Zhu et al., 2021), KET-QA (Hu et al., 2024), FinQA, TabMWP (Lu et al., 2023b) (Chen et al., 2021), and DynaQA (Lu et al., 2022) omit visual information entirely, while early multimodal efforts—MMCoQA (Li et al., 2022), MMTab…”

“While some efforts have been made to incorporate textual information as external knowledge (Chen et al., 2020; Zhu et al., 2021a; Chen et al., 2021b), there has been a significant oversight in leveraging knowledge graphs, which are widely recognized as an equally prevalent knowledge source.”

“…✓ ✓ ✗ ✗ ✗ ✗ TAT-QA (Zhuetal.,2021) ✗ ✓ ✓ ✗ ✓ ✓ ✗ MultimodelQA(Talmoretal.,2021) ✗ ✓ ✓ ✗ ✓ ✗ ✗ Manymodelqa (Hannanetal.,2020) ✗ ✓ ✓ ✗ ✓ ✗ ✗ FinQA (Chenetal.,2021b) ✗ ✓ ✓ ✗ ✓ ✓ ✗ HetpQA (Shenetal.,2022) ✗ ✓ ✓ ✗ ✓ ✗ ✗ CompMix (Christmannetal.,2023) ✗ ✓ ✓ ✓ ✓ ✗ ✗ WikiMovies-10K (Milleretal.,2016)…”

“FinQA follows the question answering system introduced by Chen et al. (2021).”

“Besides, there are many other types of IMRs, such as dolphin languages (Huang et al. 2018), domain-specific languages (Chen et al. 2021b) and equations (Roy, Upadhyay, and Roth 2016; He-Yueya et al. 2023).”

“However, MultiHiertt is the more challenging compared to the other benchmarks (Zhao et al., 2022), which does not involve either multiple tables and longer texts (Zhu et al., 2021; Chen et al., 2021; Li et al., 2022) or multi-step numerical reasoning (Chen et al., 2020b; Li et al., 2021).”

“Traditional NLP tasks in finance, such as information extraction (Chen et al., 2021a), question answering (Chen et al., 2021b; Liu et al., 2023), and summarization (Huang et al., 2024), have focused on factual accuracy.”

“Existing domain-specific benchmarks, such as FinQA [2] and FinSearchComp [7] for finance or MedQA [8] for medicine, focus on static knowledge retrieval and reasoning over historical documents.”

“Datasets like FinQA [6] and TAT-QA [30] pushed models to comprehend semi-structured data combining text and tables.”

“• OTT-QA: Open domain table-and-text question answering dataset requiring retrieval and fusion of both tabular data and textual passages from large pools.”

“Previous research has introduced benchmarks for evaluating LLMs’ abilities in various domains, including reading comprehension 40,41 , mathematical abilities 42 , reasoning about physical common sense 43 , numerical abilities over financial data 44 , and common-sense reasoning 45,46 .”

“We construct WildStruct -style fine-tuning sets from the FinQA [3] and TAT-QA [38] training corpora, yielding WMMFinQA and WMMTatQA , tailored for neuro-symbolic annealing with progressive degradation curricula.”

“For table understanding, we included datasets like TabMWP [105], TAT-QA [106], HiTab [107], MultiHiertt [108], FinQA [109], WikiSQL [110], SQA [111], and WTQ [112], contributing significantly in terms of tables, related Q/A pairs, and millions of tokens.”

“…et al., 2020]; question answering, where LLMs are tasked with answering an economic question based on the provided information [Maia et al., 2018, Chen et al., 2021, 2022, Shah et al., 2022, Xie et al., 2023b, Raman et al., 2024]; financial text summarization, which entails condensing long…”

“(LLMs) TV [65] BDD [65] AlpacaFarm [66] FinQA [67] Avg.”

“More rigorous approaches leverage external evaluation mechanisms, such as tool-interactive learning (Gou et al., 2024; Chen et al., 2021), or external knowledge sources like Wikipedia and the Internet (Xu et al., 2023; Li et al., 2024a).”

“Some works propose to evaluate models on such tasks, by presenting challenge datasets based on financial data [11, 57].”

“There are multiple datasets specifically catered for applications of NLP in finance including question answering dataset created by Chen et al. (2021) and Maia et al. (2018), and also a NER dataset constructed by Shah et al. (2023b) for the financial domain.”

“The remaining three are foundational financial question-answering references—the FinQA [15] and TAT-QA [144] benchmarks and the hierarchical-retrieval system of Choe et al. [22]—cited in the finance case study (Section 7.4) to ground that application in independent, peer-reviewed evidence; they fall…”

“FinQA [7] and TAT-QA [8] focus on numerical and tabular reasoning over financial reports.”

“Three unrelated supervised tasks are presented in a fixed order — ToolUse [50], FinQA [5], SciKE-Bio [12] with each phase training a Qwen-3-8B on a 500-example subsample of one task’s training set and validating against all three tasks at every step.”

“Existing finance benchmarks provide many of the domain ingredients for this agenda: numerical reasoning over a pre-selected page Chen et al. [2021], Zhu et al. [2021], Reddy et al. [2024], Zhao et al. [2024], retrieval ranking over filings Choi et al. [2025b,a], single-fact lookup against documents…”

“The FinanceBench end-to-end HalRate (4.9%, 1.3pp higher than FinQA) is consistent with this OOD degradation partially propagating through the pipeline, with dampening from the multi-component structure.”

“This reduces ambiguity compared to free-form textual memory and supports controlled operand reuse, denominator consistency tracking, and explicit entity disambiguation [9].”

“Tabular data is one of the most ubiquitous formats in the real world, with applications spanning finance (Chen et al., 2021; Zhu et al., 2021; 2024b), healthcare (Shi et al., 2024; Wang et al., 2024), and many other domains that rely on relational databases and spreadsheets (Fey et al., 2024).”

“However, this method faces challenges, including high costs, particularly for tasks demanding specialized expertise, such as medical (Johnson et al., 2023; Li et al., 2022) and financial (Chen et al., 2021) tasks, and is labor-intensive for large-scale annotation.”

“In finance, datasets such as FinQA (Chen et al., 2021a) and TAT-QA (Zhu et al., 2021) evaluate multi-step numerical reasoning over financial reports, Benchmarks like BizBench (Koncel-Kedziorski et al., 2023) and FinanceBench (Islam et al., 2023) further extend evaluation to programmatic reasoning and factual QA.”

“Table is a common data format in many down-stream applications (Pasupat and Liang, 2015; Parikh et al., 2020) and domains (Chen et al., 2021; Lu et al., 2023).”

“Extensive experiment analysis shows the superiority of our approach : for in-domain setting, TabAF achieves new SOTA performance on WikiTableQuestion, HiTab, and TabFact (§5.2.1); for out-of-domain setting, TabAF demonstrate generalization capabilities on AIT-QA for aviation-field tableQA and FinQA for finance-field H…”

“For example, it can not be augmented with new phrases for better adapting to an unseen domain (Koehn and Knowles, 2017; Jin et al., 2020; Chen et al., 2022) or verbatim reference text spans for better inline evidence generation (Menick et al., 2022; Gao et al., 2023).”

“Notably, Chen et al. (2022) introduce FinQA, a dataset constructed from financial statements for assessing LLMs’ multi-step numerical reasoning.”

“FinQA (Chen et al., 2021) includes QA pairs that require quantitative reasoning (with a numerical answer) over a medium-length text and a table extracted from a financial report.”

“Automated table generation has broad applications in fields such as healthcare, finance, scientific research and education (Chen et al., 2021; Johnson et al., 2016; Berant et al., 2018) where converting unstructured factual data into structured tables can significantly enhance decision-making,…”

“…including table question answering (QA) (Pasupat and Liang, 2015; Chen et al., 2020c; Nan et al., 2022; Cheng et al., 2022; Wu et al., 2025), table fact verification (Chen et al., 2020b; Aly et al., 2021), table-to-text generation (Parikh et al., 2020; Chen et al., 2020a; Nan et al., 2021), etc.”

“TAT-QA [24], FinQA [3] and Multihiertt [22] are the numerical reasoning hybrid dataset which comes from the financial field.”

“Details for each evaluation metric is available in Appendix H.”

“• FinQA (Chen et al., 2021): Released under the MIT License.”

“The 10 RAGBench subsets span five broad domains: medical (CovidQA), legal (CUAD), technical support (TechQA, EManual, DelucionQA), financial (FinQA, TatQA), and open-domain (HAGRID, Ex-pertQA, ClapNQ), with corpus sizes ranging from 221 to 12,727 documents per subset.”

“…toward numerical reasoning and question answering over financial reports, with benchmarks such as FinQA, TAT-QA, and conversational extensions (Chen et al., 2021; Zhu et al., 2021; Chen et al., 2022; Xie et al., 2025; Qian et al., 2025b), as well as retrieval-centric long-form settings, e.g.,…”

“…challenge naturally aligns with methodologies in Visual Question Answering (VQA), where models are designed to perform image-grounded reasoning by interpreting visual content in response to textual queries and producing semantically rich answers (Chen et al. 2021; Lu et al. 2022; Li et al. 2022).”

“Financial NLP benchmarks have mainly focused on knowledge retrieval, fact-based QA, and numerical reasoning over financial reports and tables (Chen et al., 2022; Zhu et al., 2021).”

“Dataset Construction: Following the methodology established by Srinivasan et al. (2025), we construct our evaluation dataset using a targeted subset of companies from FinanceBench (Islam et al. 2023) and ConvFinQA (Chen et al. 2022).”