---
title: 'TASERTab: Financial Holdings Table Dataset'
url: https://www.emergentmind.com/topics/tasertab
type: topic
---

# TASERTab: Financial Holdings Table Dataset

TASERTab is a public dataset of real-world Financial Holdings Tables released alongside TASER, a schema-guided, agentic extraction system for financial documents. It is intended for research on table detection, table extraction, normalized holdings reconstruction, and iterative schema refinement in settings where tables are multi-page, fragmented, heterogeneous, and frequently lack explicit visual boundaries. The associated paper presents TASERTab as one of the first large-scale real financial table datasets and as a benchmark specifically targeted at the difficulties of regulatory-scale financial filings, where essential holdings information may be distributed across many pages and many financial instrument types [2508.13404].

## 1. Origin, scope, and intended research use

TASERTab was introduced in the context of TASER, short for *Table Agents for Schema-guided Extraction and Recommendation*. Within that framework, the dataset serves as both training and evaluation substrate for a pipeline that performs table detection, classification, extraction, and schema recommendation from real-world financial filings. The task is not limited to locating tables: it also includes recovering holdings into normalized, schema-conforming outputs for instrument classes such as Equity, Debt, Option, Swap, Forward, Futures, ELNs, and an “Other” bucket for novelty categories [2508.13404].

The corpus is explicitly oriented toward **Financial Holdings Tables** rather than generic scientific, web, or business tables. This domain focus matters because the paper characterizes the source documents as unusually difficult: the tables are often messy, fragmented, multi-page, and semantically heterogeneous. A central design premise is therefore that a useful benchmark in this area must expose both structural difficulty and semantic open-world variability.

TASERTab is also positioned as a released community resource. The paper states that the dataset is public and includes labeled holdings tables, net asset values for each fund, extraction annotations for all holdings, and URLs to source documents. Full schema definitions, prompts, extraction metrics, and Pydantic-based validation code are part of the released benchmark package.

## 2. Corpus construction and annotation protocol

The dataset was built from public fund websites and spans a wide spectrum of fund entities globally. Its reported scale is **22,584 manually labeled pages**, **28,150,449 tokens**, and **3,213** real-world Financial Holdings Tables, with annotated investments totaling **\$731,685,511,687**. The paper describes this as manual annotation at multiple levels—page, table, and holding—which makes the corpus suitable for experiments ranging from page-level detection to holding-level normalization [2508.13404].

Annotation is not restricted to table boundaries. The released metadata includes net asset values, page spans, fund names, and source URLs. The summary also states that segmented table regions are available. This makes TASERTab relevant for evaluation protocols that need to connect document structure with downstream financial reconstruction.

Preprocessing is described as a hybrid pipeline using **pdfplumber**, OCR when needed, text normalization, and spatial metadata preservation. The preservation of spatial metadata is significant because the documents are structurally irregular; a purely text-only representation would not capture all layout cues, while a purely vision-centric representation would miss schema-level semantics and financial normalization targets.

The annotations cover major financial instrument categories: **Equities, Debt, Forwards, Swaps, Options, Futures, ELNs**, plus an **“Other”** category. That taxonomy aligns the dataset with the schema-driven extraction problem studied in TASER and supports evaluation of open-world schema extension.

## 3. Structural complexity and document topology

TASERTab is defined by structural difficulty. The paper reports that **99.4% of tables** in the dataset have **no bounding boxes**, and that the **maximum number of rows** reaches **426 per table across 44 pages**. It also reports **over 3200 unique document topologies**, with **57.53% hierarchical** tables involving spanning cells. These properties make the corpus materially different from benchmarks built around clean, single-page, explicitly delineated tables [2508.13404].

The comparison reported in the paper against PubTables-1M highlights the scale disparity:

| Property | TASERTab | PubTables-1M |
|---|---:|---:|
| Avg #Rows | 53.7 | 13.4 |
| Avg #Cols | 6.36 | 5.46 |
| Max Row Count | 426 | — |
| Max Page Span | 44 | — |

The summary reports the proportion of multi-page tables in two slightly different forms: **60.2%** multi-page in the comparison table, and elsewhere **61% multi-page** with **39% single-page**. Regardless of that reporting difference, the central point is consistent: multi-page structure is typical rather than exceptional in TASERTab.

Tables are further described as being interleaved with running text, footnotes, and images. This suggests that the benchmark stresses not only cell detection or table boundary identification, but also long-range continuity recovery across pages and robust discrimination between holdings content and surrounding document matter.

## 4. Schema, financial instrument modeling, and extraction targets

A defining feature of TASERTab is that it is coupled to a schema-guided extraction setting rather than a generic table-parsing task. In TASER, the active schema is implemented with **Pydantic models**, using a modular structure built around a base `Instrument` model and subclasses for specific instrument types. The paper’s excerpted `Portfolio` model includes fields such as `fund_name`, `value_in_thousands`, and instrument-specific lists such as `equities`, with analogous fields for the remaining instrument categories [2508.13404].

This schema orientation means that the dataset’s target output is not merely a reconstructed table image or grid; it is a set of strongly typed holdings entries validated against the schema. In the TASER pipeline, non-conforming entries are treated as unmatched holdings and become inputs to later schema refinement. That design links TASERTab to a broader open-world extraction problem: financial documents may contain legitimate holdings that are absent from the initial schema.

The paper formalizes the iterative refinement setting using unmatched holdings $H=\{h_1,\ldots,h_N\}$ and an initial schema $\Sigma^{(0)}$. At iteration $\ell$, unmatched holdings are partitioned into batches of size $B$, schema suggestions are proposed for each batch, suggestions are aggregated and clustered, the schema is updated to $\Sigma^{(\ell+1)}$, and extraction is rerun until unmatched holdings are exhausted or no new suggestions remain. As a dataset, TASERTab therefore supports experiments not only in detection and extraction, but also in schema evolution under recursive validation.

A plausible implication is that TASERTab is unusually well suited to evaluating systems that combine layout understanding with typed information extraction, especially when the instrument ontology is incomplete at initialization.

## 5. Benchmarking role and reported empirical results

The TASER paper uses TASERTab as the evaluation bed for detection, extraction, and schema-refinement studies. The reported metrics include **Recall, Precision, F1, Accuracy** for detection; **Total Absolute Difference (TAD = |sum_extractions - ground_truth|)** and **Percent Unaccounted Holdings Value** for extraction; and **Coverage, Diversity (pairwise Levenshtein),** and **Collision rates** for schema refinement [2508.13404].

On table detection, the paper reports a **10.1% improvement** over Table Transformer. The specific comparison given is **precision = 43.43%** for TASER “Full Schema Prompting” versus **32.75%** for Table Transformer. The same summary states that TASER achieves **100% recall** for holdings tables in all ablations and the highest F1 and accuracy across the ablation strategies.

For extraction quality, the paper reports **\$102.8M TAD** for TASER, compared with **\$107.1–\$120.6M** for baselines, together with **0.014% unaccounted value** against **\$731.7B** labeled holdings. These figures indicate that the benchmark is evaluated not only by structural correctness but also by downstream financial value recovery.

The schema-learning experiments show that batch size materially affects performance. The paper reports that larger batch sizes yield a **104.3% increase in actionable schema recommendations that are actionable and utilized**, resulting in a **9.8% increase in extracted holdings**. It also reports a trade-off between **coverage** and **utilization**: for **batch size 10**, **coverage = 96.1%**, **utilization = 29.0%**, **collisions = 73.5%**; for **batch size 500**, **coverage = 36.0%**, **utilization = 59.2%**, **collisions = 14.0%**. The same analysis states that resolving the largest unmatched holdings reduces the absolute dollar difference by **7–10%** across batch sizes, with most value recovered by reconciling only the top **10–20%** of holdings.

## 6. Methodological significance and common misconceptions

TASERTab is important partly because it challenges a common assumption that table extraction benchmarks are primarily about visually clean, bounded, single-page tables. The dataset’s reported properties—**99.4%** without bounding boxes, **426** maximum rows, spans up to **44 pages**, and a majority of multi-page examples—indicate a regime in which conventional table assumptions are systematically violated [2508.13404].

Another misconception is that financial table extraction in this setting is only a detection problem. TASERTab, as released with TASER, is tied to evaluation of page-level localization, holding-level extraction, schema conformance, and financial value recovery. The benchmark therefore couples document understanding with semantic normalization and ontology management.

The paper’s characterization of TASERTab as the **first large-scale, multi-instrument, real-world annotated corpus** for financial tables should be read in the context of its stated goal: enabling benchmarking for robust, agentic table understanding in real regulatory documents. This suggests a specific research agenda rather than a generic tabular benchmark. TASERTab is domain-specific, but its methodological relevance extends to any extraction problem in which records are distributed across long documents, expressed in heterogeneous layouts, and only partially covered by an initial schema.

In that sense, TASERTab occupies a distinctive position among released datasets: it is simultaneously a corpus of difficult financial tables, a benchmark for schema-guided extraction, and a testbed for continuous learning through recursive schema refinement.

Source: https://www.emergentmind.com/topics/tasertab