Blame for README.md

---
license: cc-by-nc-4.0
task_categories:
- question-answering
- text-generation
language:
- en
tags:
- finance
- knowledge-graphs
- multi-hop-reasoning
- multi-hop-qa
- sec-filings
- 10-k-filings
- financial-nlp
- document-grounding
- benchmark
- FinReflectKG
pretty_name: FinMultiHopQA
size_categories:
- 10K<n<100K
configs:
- config_name: default
  data_files:
  - split: test
    path: data/test-*
---

# Dataset Card for FinMultiHopQA

**FinMultiHopQA** is a KG-grounded benchmark of **15,600 qualitative and quantitative multi-hop questions**
over SEC 10-K filings. Every question is built from a typed path in **FinReflectKG**, a provenance-linked
financial knowledge graph, and ships with the exact evidence chunks (filing, page, chunk) for each hop.

## Dataset Details

### Dataset Description

**Curated by:** [Domyn](https://www.domyn.com/)

**Language(s):** English

**License:** CC-BY-NC-4.0

FinMultiHopQA tests whether models can combine evidence that is scattered across pages, filings, fiscal
years, and companies. Questions require 2 or 3 reasoning hops and cover three document regimes:
intra-document, inter-year (same company, different fiscal years), and cross-company. Answers are grounded
only in the released evidence chunks, so no external knowledge is needed.

### Dataset Sources

**Base knowledge graph:** [FinReflectKG](https://huggingface.co/datasets/iGeniusAI/FinReflectKG) |
**Paper:** [FinReflectKG: Agentic Construction and Evaluation of Financial Knowledge Graphs](https://arxiv.org/pdf/2508.17906)

**Benchmark paper:** FinMultiHopQA: Financial QA Benchmark for Reasoning with Knowledge Graph Evidence

## Key Features

- **15,600 QA pairs**: 7,800 qualitative and 7,800 quantitative; 6,000 2-hop and 9,600 3-hop
- **Three document regimes**: intra-document, inter-year, and cross-company
- **Two 3-hop motif families**: connector chains (`type1`) and three-anchor branches (`type2`)
- **Full provenance**: every hop carries its source filing, page, chunk ID, and chunk text
- **Coverage**: 96 companies, 267 10-K filings, fiscal years 2022–2024, 11 GICS sectors
- **Quality-screened**: every pair scored by an LLM validator (≥ 45/50) plus hard rejection rules
- **Expert-validated**: a two-SME quality audit on a 360-pair random sample

![Inter-year multi-hop instance: KG path, per-hop evidence chunks, and reference answer](assets/kg_multihop_inter_year_1.png)

## Quick Start

```python
from datasets import load_dataset

ds = load_dataset("domyn/FinMultiHopQA", split="test")  # 15,600 questions

# Filter by question type, hop count, topology, or motif
quant = ds.filter(lambda r: r["question_type"] == "quantitative")
three_hop = ds.filter(lambda r: r["hop_count"] == 3)
inter_year = ds.filter(lambda r: r["topology"] == "cross_year" and r["motif"] == "chain")

row = ds[0]
print(row["question"])
for hop in row["hops"]:
    print(hop["hop"], hop["relationship"], hop["source_file"], hop["page_id"])
```

## Dataset Structure

### Files and splits

The full benchmark is a single `test` split (15,600 rows). Use the `question_type`, `hop_count`,
`topology`, and `motif` columns to select subsets.

### Statistics

| Question type | Topology | 2-hop | 3-hop | Total |
|---|---|---:|---:|---:|
| Qualitative | Cross-company | 1,250 | 2,000 | 3,250 |
| Qualitative | Inter-year | 1,000 | 1,600 | 2,600 |
| Qualitative | Intra-document | 750 | 1,200 | 1,950 |
| Quantitative | Cross-company | 1,250 | 2,000 | 3,250 |
| Quantitative | Inter-year | 1,000 | 1,600 | 2,600 |
| Quantitative | Intra-document | 750 | 1,200 | 1,950 |
| **Total** | | **6,000** | **9,600** | **15,600** |

Motifs: 10,000 chain (all 2-hop + 3-hop `type1`) and 5,600 branch (3-hop `type2`).

![Hop count and document topology across the benchmark](assets/hop_count_doc_type_topology.png)

**Sector coverage (GICS):** Information Technology 2,321 · Industrials 2,106 · Health Care 2,019 ·
Financials 1,952 · Consumer Staples 1,828 · Consumer Discretionary 1,611 · Communication Services 1,518 ·
Energy 853 · Utilities 853 · Real Estate 371 · Materials 168

![Sector distribution, full generation corpus versus filtered benchmark](assets/sector_distribution_full_vs_filtered.png)

### Partitions

Each partition keeps the top-N candidates by `quality_score` (descending), then source question ID (ascending).

| Partition | Questions | Sampling | Quality score (min / max / mean) |
|---|---:|---|---|
| `qualitative_2hop_cross_company` | 1,250 | `top_quality_score` | 46 / 49 / 46.78 |
| `qualitative_2hop_cross_year` | 1,000 | `top_quality_score` | 46 / 49 / 46.64 |
| `qualitative_2hop_intra_doc` | 750 | `top_quality_score` | 46 / 50 / 47.14 |
| `qualitative_3hop_cross_company_type1` | 600 | `top_quality_score` | 47 / 48 / 47.03 |
| `qualitative_3hop_cross_company_type2` | 1,400 | `top_quality_score` | 47 / 48 / 47.15 |
| `qualitative_3hop_cross_year_type1` | 800 | `top_quality_score` | 47 / 48 / 47.02 |
| `qualitative_3hop_cross_year_type2` | 800 | `top_quality_score` | 46 / 48 / 47.06 |
| `qualitative_3hop_intra_doc_type1` | 600 | `top_quality_score` | 47 / 48 / 47.03 |
| `qualitative_3hop_intra_doc_type2` | 600 | `top_quality_score` | 46 / 48 / 46.72 |
| `quantitative_2hop_cross_company` | 1,250 | `top_quality_score` | 47 / 49 / 47.13 |
| `quantitative_2hop_cross_year` | 1,000 | `top_quality_score` | 47 / 49 / 47.08 |
| `quantitative_2hop_intra_doc` | 750 | `top_quality_score` | 47 / 50 / 47.27 |
| `quantitative_3hop_cross_company_type1` | 600 | `top_quality_score` | 46 / 48 / 46.97 |
| `quantitative_3hop_cross_company_type2` | 1,400 | `top_quality_score` | 47 / 48 / 47.09 |
| `quantitative_3hop_cross_year_type1` | 800 | `top_quality_score` | 47 / 48 / 47.02 |
| `quantitative_3hop_cross_year_type2` | 800 | `top_quality_score` | 47 / 49 / 47.22 |
| `quantitative_3hop_intra_doc_type1` | 600 | `top_quality_score` | 47 / 48 / 47.03 |
| `quantitative_3hop_intra_doc_type2` | 600 | `top_quality_score` | 46 / 48 / 46.82 |

### Data Fields

| Field | Type | Description |
|---|---|---|
| `id` | string | Unique key: `<partition>-<row index>` |
| `question` | string | The multi-hop question |
| `answer` | string | Reference answer grounded in the evidence chunks |
| `reasoning_steps` | list[string] | Step-by-step reasoning chain behind the reference answer |
| `question_type` | string | `qualitative` or `quantitative` |
| `hop_count` | int | `2` or `3` |
| `topology` | string | `intra_doc`, `cross_year`, or `cross_company` (from the partition; use this for topology analysis) |
| `motif` | string | `chain` (2-hop and 3-hop `type1`) or `branch` (3-hop `type2`) |
| `pattern` | string | Typed KG path, e.g. `ORG -[Invests_In]-> GPE <-[Operates_In]- ORG` |
| `document_relationship` | string | `intra_document`, `inter_document_same_company`, or `inter_document_cross_company` |
| `sector` | string | GICS sector |
| `sub_industry` | string | GICS sub-industry (null when not resolved) |
| `difficulty` | string | IDF-based heuristic: `easy` / `medium` / `hard` |
| `idf_score` | float | Connector IDF score (higher = more specific connector) |
| `quality_score` | int | LLM-validator rubric total (0–50) |
| `nodes` | list[struct] | KG path nodes: `{role, id, name, type, year, idf_score}` |
| `hops` | list[struct] | Evidence per hop: `{hop, relationship, source_file, page_id, chunk_id, chunk_text}` |

Node roles: 2-hop paths use `start`, `intermediate`, `end`; 3-hop chains use `start`, `node_2`,
`node_3`, `end`; 3-hop branches use `start`, `connector` (the shared node the anchors link to), `end`.
For every row, `hops[i].source_file` gives the filing behind hop *i*.

### Example

```json
{
  "id": "qualitative_3hop_cross_company_type2-00000",
  "question": "How do Boeing (BA), Emerson (EMR), and Honeywell (HON) each approach the strategic asset allocation for their defined benefit pension plans, ...?",
  "answer": "Boeing focuses on reducing asset volatility relative to liabilities, ...",
  "reasoning_steps": ["Hop 1: BA → Asset Allocation: ...", "Hop 2: EMR → Asset Allocation: ...", "Hop 3: HON → Asset Allocation: ..."],
  "question_type": "qualitative",
  "hop_count": 3,
  "topology": "cross_company",
  "motif": "branch",
  "pattern": "ORG -[Monitors]-> ACCOUNTING_POLICY <-[Discloses]- ORG <-[Discloses]- ORG",
  "sector": "Industrials",
  "difficulty": "medium",
  "quality_score": 48,
  "nodes": [
    {"role": "start", "id": "BA", "name": "BA", "type": "ORG", "year": 2024, "idf_score": null},
    {"role": "connector", "id": "Asset_Allocation", "name": "Asset Allocation", "type": "ACCOUNTING_POLICY", "year": null, "idf_score": 4.01},
    {"role": "end", "id": null, "name": "EMR", "type": "ORG", "year": null, "idf_score": null}
  ],
  "hops": [
    {"hop": 1, "relationship": "Monitors", "source_file": "BA_10k_2024.pdf", "page_id": "page_99", "chunk_id": "chunk_1", "chunk_text": "..."},
    {"hop": 2, "relationship": "Discloses", "source_file": "EMR_10k_2024.pdf", "page_id": "page_63", "chunk_id": "chunk_5", "chunk_text": "..."},
    {"hop": 3, "relationship": "Discloses", "source_file": "HON_10k_2024.pdf", "page_id": "page_107", "chunk_id": "chunk_3", "chunk_text": "..."}
  ]
}
```

## Dataset Creation

### Motif families

Questions are instantiated from typed KG motifs, each under three evidence regimes with explicit
provenance constraints:

- **2-hop bridge** (`cross_company`, `cross_year`, `intra_doc`): cross-company uses two distinct
  organizations with matched fiscal years; inter-year uses one organization across different years;
  intra-document uses one document but distinct pages/chunks (or distinct facts in a shared chunk).
- **3-hop connector chain** (`type1`): connector pairs are chosen by IDF ranking within GICS strata, which
  suppresses generic connectors; the linking relation must share provenance with the anchor edges.
- **3-hop multi-branch** (`type2`): three organization anchors (or one organization across three years)
  each linked independently to one shared connector node.

### Generation and filtering

The pipeline generates a full corpus of **36,668** QA pairs; this benchmark is the filtered subset of
**15,600**. Filtering preserves the topology and sector balance of the full corpus.

### Quality control

Every generated pair is scored by an LLM validator on five dimensions (0–10 each, max 50):
multi-chunk dependency, financial context, answer grounding, causal-connection grounding, and natural
analyst quality. Pairs below **45/50** are regenerated. Independent of the score, a pair is discarded if
it:

- mentions page or section structure;
- is answerable from one evidence fact without combining the required hops;
- needs external knowledge not present in the evidence;
- asserts a fabricated causal link between companies that only share a generic term.

The rubric total is released as `quality_score` on every row.

![13 schema-aware motif categories](assets/motif_distribution.png)

### Human (SME) validation

Two finance SMEs independently rated a fixed-seed, topology-stratified random sample of 360 QA pairs on
four criteria. A separate SME judge-calibration study scored model answers on a distinct 360-pair panel
(1,080 assessments).

## Uses

- Evaluating LLM multi-hop reasoning over financial filings
- Comparing evidence-delivery strategies (KG-linked chunks vs. page windows vs. retrieval)
- Studying how performance changes with hop count, topology, and motif

The benchmark paper evaluates six evidence protocols: KG-linked provenance (E1), page windows (E2), page
windows with distractors (E3), relevance-sorted descending (E4) and ascending (E5), and semantic retrieval
(E6).

## Known Issues

- **`difficulty`** is an IDF heuristic (~80% `hard`), not a calibrated difficulty measure.

## Citation

If you use this dataset, please cite:

```bibtex
@article{arun2026finmultihopqa,
  title={FinMultiHopQA: Financial QA Benchmark for Reasoning with Knowledge Graph Evidence},
  author={Arun, Abhinav and Harsh, Reetu Raj and Sarmah, Bhaskarjit and Pasquali, Stefano},
  journal={arXiv preprint},
  year={2026}
}
```

FinMultiHopQA is built on FinReflectKG:

```bibtex
@article{arun2025finreflectkg,
  title={FinReflectKG: Agentic Construction and Evaluation of Financial Knowledge Graphs},
  author={Arun, Abhinav and Dimino, Fabrizio and Agarwal, Tejas Prakash and Sarmah, Bhaskarjit and Pasquali, Stefano},
  journal={arXiv preprint arXiv:2508.17906},
  year={2025},
  url={https://arxiv.org/pdf/2508.17906}
}
```