Datasets:
Add FinMultiHopQA benchmark (15,600 QA pairs) with dataset card
Browse files- README.md +370 -0
- assets/hop_count_doc_type_topology.png +3 -0
- assets/kg_multihop_inter_year_1.png +3 -0
- assets/motif_distribution.png +3 -0
- assets/sector_distribution_full_vs_filtered.png +3 -0
- data/qualitative_2hop_cross_company.parquet +3 -0
- data/qualitative_2hop_cross_year.parquet +3 -0
- data/qualitative_2hop_intra_doc.parquet +3 -0
- data/qualitative_3hop_cross_company_type1.parquet +3 -0
- data/qualitative_3hop_cross_company_type2.parquet +3 -0
- data/qualitative_3hop_cross_year_type1.parquet +3 -0
- data/qualitative_3hop_cross_year_type2.parquet +3 -0
- data/qualitative_3hop_intra_doc_type1.parquet +3 -0
- data/qualitative_3hop_intra_doc_type2.parquet +3 -0
- data/quantitative_2hop_cross_company.parquet +3 -0
- data/quantitative_2hop_cross_year.parquet +3 -0
- data/quantitative_2hop_intra_doc.parquet +3 -0
- data/quantitative_3hop_cross_company_type1.parquet +3 -0
- data/quantitative_3hop_cross_company_type2.parquet +3 -0
- data/quantitative_3hop_cross_year_type1.parquet +3 -0
- data/quantitative_3hop_cross_year_type2.parquet +3 -0
- data/quantitative_3hop_intra_doc_type1.parquet +3 -0
- data/quantitative_3hop_intra_doc_type2.parquet +3 -0
README.md
ADDED
|
@@ -0,0 +1,370 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
task_categories:
|
| 3 |
+
- question-answering
|
| 4 |
+
- text-generation
|
| 5 |
+
language:
|
| 6 |
+
- en
|
| 7 |
+
tags:
|
| 8 |
+
- finance
|
| 9 |
+
- knowledge-graphs
|
| 10 |
+
- multi-hop-reasoning
|
| 11 |
+
- multi-hop-qa
|
| 12 |
+
- sec-filings
|
| 13 |
+
- 10-k-filings
|
| 14 |
+
- financial-nlp
|
| 15 |
+
- document-grounding
|
| 16 |
+
- benchmark
|
| 17 |
+
- FinReflectKG
|
| 18 |
+
pretty_name: FinMultiHopQA
|
| 19 |
+
size_categories:
|
| 20 |
+
- 10K<n<100K
|
| 21 |
+
configs:
|
| 22 |
+
- config_name: default
|
| 23 |
+
default: true
|
| 24 |
+
data_files:
|
| 25 |
+
- split: test
|
| 26 |
+
path: data/*.parquet
|
| 27 |
+
- config_name: qualitative
|
| 28 |
+
data_files:
|
| 29 |
+
- split: test
|
| 30 |
+
path: data/qualitative_*.parquet
|
| 31 |
+
- config_name: quantitative
|
| 32 |
+
data_files:
|
| 33 |
+
- split: test
|
| 34 |
+
path: data/quantitative_*.parquet
|
| 35 |
+
- config_name: qualitative_2hop_cross_company
|
| 36 |
+
data_files:
|
| 37 |
+
- split: test
|
| 38 |
+
path: data/qualitative_2hop_cross_company.parquet
|
| 39 |
+
- config_name: qualitative_2hop_cross_year
|
| 40 |
+
data_files:
|
| 41 |
+
- split: test
|
| 42 |
+
path: data/qualitative_2hop_cross_year.parquet
|
| 43 |
+
- config_name: qualitative_2hop_intra_doc
|
| 44 |
+
data_files:
|
| 45 |
+
- split: test
|
| 46 |
+
path: data/qualitative_2hop_intra_doc.parquet
|
| 47 |
+
- config_name: qualitative_3hop_cross_company_type1
|
| 48 |
+
data_files:
|
| 49 |
+
- split: test
|
| 50 |
+
path: data/qualitative_3hop_cross_company_type1.parquet
|
| 51 |
+
- config_name: qualitative_3hop_cross_company_type2
|
| 52 |
+
data_files:
|
| 53 |
+
- split: test
|
| 54 |
+
path: data/qualitative_3hop_cross_company_type2.parquet
|
| 55 |
+
- config_name: qualitative_3hop_cross_year_type1
|
| 56 |
+
data_files:
|
| 57 |
+
- split: test
|
| 58 |
+
path: data/qualitative_3hop_cross_year_type1.parquet
|
| 59 |
+
- config_name: qualitative_3hop_cross_year_type2
|
| 60 |
+
data_files:
|
| 61 |
+
- split: test
|
| 62 |
+
path: data/qualitative_3hop_cross_year_type2.parquet
|
| 63 |
+
- config_name: qualitative_3hop_intra_doc_type1
|
| 64 |
+
data_files:
|
| 65 |
+
- split: test
|
| 66 |
+
path: data/qualitative_3hop_intra_doc_type1.parquet
|
| 67 |
+
- config_name: qualitative_3hop_intra_doc_type2
|
| 68 |
+
data_files:
|
| 69 |
+
- split: test
|
| 70 |
+
path: data/qualitative_3hop_intra_doc_type2.parquet
|
| 71 |
+
- config_name: quantitative_2hop_cross_company
|
| 72 |
+
data_files:
|
| 73 |
+
- split: test
|
| 74 |
+
path: data/quantitative_2hop_cross_company.parquet
|
| 75 |
+
- config_name: quantitative_2hop_cross_year
|
| 76 |
+
data_files:
|
| 77 |
+
- split: test
|
| 78 |
+
path: data/quantitative_2hop_cross_year.parquet
|
| 79 |
+
- config_name: quantitative_2hop_intra_doc
|
| 80 |
+
data_files:
|
| 81 |
+
- split: test
|
| 82 |
+
path: data/quantitative_2hop_intra_doc.parquet
|
| 83 |
+
- config_name: quantitative_3hop_cross_company_type1
|
| 84 |
+
data_files:
|
| 85 |
+
- split: test
|
| 86 |
+
path: data/quantitative_3hop_cross_company_type1.parquet
|
| 87 |
+
- config_name: quantitative_3hop_cross_company_type2
|
| 88 |
+
data_files:
|
| 89 |
+
- split: test
|
| 90 |
+
path: data/quantitative_3hop_cross_company_type2.parquet
|
| 91 |
+
- config_name: quantitative_3hop_cross_year_type1
|
| 92 |
+
data_files:
|
| 93 |
+
- split: test
|
| 94 |
+
path: data/quantitative_3hop_cross_year_type1.parquet
|
| 95 |
+
- config_name: quantitative_3hop_cross_year_type2
|
| 96 |
+
data_files:
|
| 97 |
+
- split: test
|
| 98 |
+
path: data/quantitative_3hop_cross_year_type2.parquet
|
| 99 |
+
- config_name: quantitative_3hop_intra_doc_type1
|
| 100 |
+
data_files:
|
| 101 |
+
- split: test
|
| 102 |
+
path: data/quantitative_3hop_intra_doc_type1.parquet
|
| 103 |
+
- config_name: quantitative_3hop_intra_doc_type2
|
| 104 |
+
data_files:
|
| 105 |
+
- split: test
|
| 106 |
+
path: data/quantitative_3hop_intra_doc_type2.parquet
|
| 107 |
+
---
|
| 108 |
+
|
| 109 |
+
# Dataset Card for FinMultiHopQA
|
| 110 |
+
|
| 111 |
+
**FinMultiHopQA** is a KG-grounded benchmark of **15,600 qualitative and quantitative multi-hop questions**
|
| 112 |
+
over SEC 10-K filings. Every question is built from a typed path in **FinReflectKG**, a provenance-linked
|
| 113 |
+
financial knowledge graph, and ships with the exact evidence chunks (filing, page, chunk) for each hop.
|
| 114 |
+
|
| 115 |
+
## Dataset Details
|
| 116 |
+
|
| 117 |
+
### Dataset Description
|
| 118 |
+
|
| 119 |
+
**Curated by:** [Domyn](https://www.domyn.com/)
|
| 120 |
+
|
| 121 |
+
**Language(s):** English
|
| 122 |
+
|
| 123 |
+
FinMultiHopQA tests whether models can combine evidence that is scattered across pages, filings, fiscal
|
| 124 |
+
years, and companies. Questions require 2 or 3 reasoning hops and cover three document regimes:
|
| 125 |
+
intra-document, inter-year (same company, different fiscal years), and cross-company. Answers are grounded
|
| 126 |
+
only in the released evidence chunks, so no external knowledge is needed.
|
| 127 |
+
|
| 128 |
+
### Dataset Sources
|
| 129 |
+
|
| 130 |
+
**Base knowledge graph:** [FinReflectKG](https://huggingface.co/datasets/iGeniusAI/FinReflectKG) |
|
| 131 |
+
**Paper:** [FinReflectKG: Agentic Construction and Evaluation of Financial Knowledge Graphs](https://arxiv.org/pdf/2508.17906)
|
| 132 |
+
|
| 133 |
+
**Benchmark paper:** FinMultiHopQA: Financial QA Benchmark for Reasoning with Knowledge Graph Evidence
|
| 134 |
+
|
| 135 |
+
## Key Features
|
| 136 |
+
|
| 137 |
+
- **15,600 QA pairs**: 7,800 qualitative and 7,800 quantitative; 6,000 2-hop and 9,600 3-hop
|
| 138 |
+
- **Three document regimes**: intra-document, inter-year, and cross-company
|
| 139 |
+
- **Two 3-hop motif families**: connector chains (`type1`) and three-anchor branches (`type2`)
|
| 140 |
+
- **Full provenance**: every hop carries its source filing, page, chunk ID, and chunk text
|
| 141 |
+
- **Coverage**: 96 companies, 267 10-K filings, fiscal years 2022–2024, 11 GICS sectors
|
| 142 |
+
- **Quality-screened**: every pair scored by an LLM validator (≥ 45/50) plus hard rejection rules
|
| 143 |
+
- **Expert-validated**: a two-SME quality audit on a 360-pair random sample
|
| 144 |
+
|
| 145 |
+

|
| 146 |
+
|
| 147 |
+
## Quick Start
|
| 148 |
+
|
| 149 |
+
```python
|
| 150 |
+
from datasets import load_dataset
|
| 151 |
+
|
| 152 |
+
# Full benchmark (15,600 questions)
|
| 153 |
+
ds = load_dataset("domyn/FinMultiHopQA", split="test")
|
| 154 |
+
|
| 155 |
+
# By question type
|
| 156 |
+
qual = load_dataset("domyn/FinMultiHopQA", "qualitative", split="test")
|
| 157 |
+
quant = load_dataset("domyn/FinMultiHopQA", "quantitative", split="test")
|
| 158 |
+
|
| 159 |
+
# A single partition
|
| 160 |
+
ds = load_dataset("domyn/FinMultiHopQA", "quantitative_3hop_cross_year_type1", split="test")
|
| 161 |
+
|
| 162 |
+
row = ds[0]
|
| 163 |
+
print(row["question"])
|
| 164 |
+
for hop in row["hops"]:
|
| 165 |
+
print(hop["hop"], hop["relationship"], hop["source_file"], hop["page_id"])
|
| 166 |
+
```
|
| 167 |
+
|
| 168 |
+
## Dataset Structure
|
| 169 |
+
|
| 170 |
+
### Configurations
|
| 171 |
+
|
| 172 |
+
All data is in a single `test` split (this is an evaluation benchmark).
|
| 173 |
+
|
| 174 |
+
| Config | Questions | Contents |
|
| 175 |
+
|---|---:|---|
|
| 176 |
+
| `default` | 15,600 | All 18 partitions |
|
| 177 |
+
| `qualitative` | 7,800 | All qualitative partitions |
|
| 178 |
+
| `quantitative` | 7,800 | All quantitative partitions |
|
| 179 |
+
| `<question_type>_<hop>_<topology>[_type1\|_type2]` | 600–1,400 | One partition (18 configs, listed below) |
|
| 180 |
+
|
| 181 |
+
### Statistics
|
| 182 |
+
|
| 183 |
+
| Question type | Topology | 2-hop | 3-hop | Total |
|
| 184 |
+
|---|---|---:|---:|---:|
|
| 185 |
+
| Qualitative | Cross-company | 1,250 | 2,000 | 3,250 |
|
| 186 |
+
| Qualitative | Inter-year | 1,000 | 1,600 | 2,600 |
|
| 187 |
+
| Qualitative | Intra-document | 750 | 1,200 | 1,950 |
|
| 188 |
+
| Quantitative | Cross-company | 1,250 | 2,000 | 3,250 |
|
| 189 |
+
| Quantitative | Inter-year | 1,000 | 1,600 | 2,600 |
|
| 190 |
+
| Quantitative | Intra-document | 750 | 1,200 | 1,950 |
|
| 191 |
+
| **Total** | | **6,000** | **9,600** | **15,600** |
|
| 192 |
+
|
| 193 |
+
Motifs: 10,000 chain (all 2-hop + 3-hop `type1`) and 5,600 branch (3-hop `type2`).
|
| 194 |
+
|
| 195 |
+

|
| 196 |
+
|
| 197 |
+
**Sector coverage (GICS):** Information Technology 2,321 · Industrials 2,106 · Health Care 2,019 ·
|
| 198 |
+
Financials 1,952 · Consumer Staples 1,828 · Consumer Discretionary 1,611 · Communication Services 1,518 ·
|
| 199 |
+
Energy 853 · Utilities 853 · Real Estate 371 · Materials 168
|
| 200 |
+
|
| 201 |
+

|
| 202 |
+
|
| 203 |
+
### Partitions
|
| 204 |
+
|
| 205 |
+
Each partition keeps the top-N candidates by `quality_score` (descending), then `question_id` (ascending).
|
| 206 |
+
|
| 207 |
+
| Partition | Questions | Sampling | Quality score (min / max / mean) |
|
| 208 |
+
|---|---:|---|---|
|
| 209 |
+
| `qualitative_2hop_cross_company` | 1,250 | `top_quality_score` | 46 / 49 / 46.78 |
|
| 210 |
+
| `qualitative_2hop_cross_year` | 1,000 | `top_quality_score` | 46 / 49 / 46.64 |
|
| 211 |
+
| `qualitative_2hop_intra_doc` | 750 | `top_quality_score` | 46 / 50 / 47.14 |
|
| 212 |
+
| `qualitative_3hop_cross_company_type1` | 600 | `top_quality_score` | 47 / 48 / 47.03 |
|
| 213 |
+
| `qualitative_3hop_cross_company_type2` | 1,400 | `top_quality_score` | 47 / 48 / 47.15 |
|
| 214 |
+
| `qualitative_3hop_cross_year_type1` | 800 | `top_quality_score` | 47 / 48 / 47.02 |
|
| 215 |
+
| `qualitative_3hop_cross_year_type2` | 800 | `top_quality_score` | 46 / 48 / 47.06 |
|
| 216 |
+
| `qualitative_3hop_intra_doc_type1` | 600 | `top_quality_score` | 47 / 48 / 47.03 |
|
| 217 |
+
| `qualitative_3hop_intra_doc_type2` | 600 | `top_quality_score` | 46 / 48 / 46.72 |
|
| 218 |
+
| `quantitative_2hop_cross_company` | 1,250 | `top_quality_score` | 47 / 49 / 47.13 |
|
| 219 |
+
| `quantitative_2hop_cross_year` | 1,000 | `top_quality_score` | 47 / 49 / 47.08 |
|
| 220 |
+
| `quantitative_2hop_intra_doc` | 750 | `top_quality_score` | 47 / 50 / 47.27 |
|
| 221 |
+
| `quantitative_3hop_cross_company_type1` | 600 | `top_quality_score` | 46 / 48 / 46.97 |
|
| 222 |
+
| `quantitative_3hop_cross_company_type2` | 1,400 | `top_quality_score` | 47 / 48 / 47.09 |
|
| 223 |
+
| `quantitative_3hop_cross_year_type1` | 800 | `top_quality_score` | 47 / 48 / 47.02 |
|
| 224 |
+
| `quantitative_3hop_cross_year_type2` | 800 | `top_quality_score` | 47 / 49 / 47.22 |
|
| 225 |
+
| `quantitative_3hop_intra_doc_type1` | 600 | `top_quality_score` | 47 / 48 / 47.03 |
|
| 226 |
+
| `quantitative_3hop_intra_doc_type2` | 600 | `top_quality_score` | 46 / 48 / 46.82 |
|
| 227 |
+
|
| 228 |
+
### Data Fields
|
| 229 |
+
|
| 230 |
+
| Field | Type | Description |
|
| 231 |
+
|---|---|---|
|
| 232 |
+
| `id` | string | Unique key: `<partition>-<row index>` |
|
| 233 |
+
| `question_id` | int | Generator ID; **not unique** (repeats across generation runs) |
|
| 234 |
+
| `question` | string | The multi-hop question |
|
| 235 |
+
| `answer` | string | Reference answer grounded in the evidence chunks |
|
| 236 |
+
| `reasoning_steps` | list[string] | Step-by-step reasoning chain behind the reference answer |
|
| 237 |
+
| `question_type` | string | `qualitative` or `quantitative` |
|
| 238 |
+
| `hop_count` | int | `2` or `3` |
|
| 239 |
+
| `topology` | string | `intra_doc`, `cross_year`, or `cross_company` (from the partition; use this for topology analysis) |
|
| 240 |
+
| `motif` | string | `chain` (2-hop and 3-hop `type1`) or `branch` (3-hop `type2`) |
|
| 241 |
+
| `partition` | string | Source partition / config name |
|
| 242 |
+
| `pattern` | string | Typed KG path, e.g. `ORG -[Invests_In]-> GPE <-[Operates_In]- ORG` |
|
| 243 |
+
| `document_relationship` | string | Original label (see Known Issues) |
|
| 244 |
+
| `sector` | string | GICS sector |
|
| 245 |
+
| `sub_industry` | string | GICS sub-industry (null when not resolved) |
|
| 246 |
+
| `difficulty` | string | IDF-based heuristic: `easy` / `medium` / `hard` |
|
| 247 |
+
| `idf_score` | float | Connector IDF score (higher = more specific connector) |
|
| 248 |
+
| `quality_score` | int | LLM-validator rubric total (0–50) |
|
| 249 |
+
| `temporal_change` | string | Present only on 2-hop inter-year rows; always `unknown` in this release |
|
| 250 |
+
| `entities` | list[struct] | Question anchors: `{role, name, type}` |
|
| 251 |
+
| `nodes` | list[struct] | KG path nodes: `{role, id, name, type, year, idf_score}` |
|
| 252 |
+
| `hops` | list[struct] | Evidence per hop: `{hop, relationship, source_file, page_id, chunk_id, chunk_text}` |
|
| 253 |
+
| `generation_file` | string | Generation-run file the pair came from |
|
| 254 |
+
|
| 255 |
+
Node roles: 2-hop paths use `start`, `intermediate`, `end`; 3-hop chains use `start`, `node_2`,
|
| 256 |
+
`node_3`, `end`; 3-hop branches use `start`, `connector` (the shared node that all three anchors link to).
|
| 257 |
+
For every row, `hops[i].source_file` gives the filing behind hop *i*.
|
| 258 |
+
|
| 259 |
+
### Example
|
| 260 |
+
|
| 261 |
+
```json
|
| 262 |
+
{
|
| 263 |
+
"id": "qualitative_3hop_cross_company_type2-00000",
|
| 264 |
+
"question": "How do Boeing (BA), Emerson (EMR), and Honeywell (HON) each approach the strategic asset allocation for their defined benefit pension plans, ...?",
|
| 265 |
+
"answer": "Boeing focuses on reducing asset volatility relative to liabilities, ...",
|
| 266 |
+
"reasoning_steps": ["Hop 1: BA → Asset Allocation: ...", "Hop 2: EMR → Asset Allocation: ...", "Hop 3: HON → Asset Allocation: ..."],
|
| 267 |
+
"question_type": "qualitative",
|
| 268 |
+
"hop_count": 3,
|
| 269 |
+
"topology": "cross_company",
|
| 270 |
+
"motif": "branch",
|
| 271 |
+
"pattern": "ORG -[Monitors]-> ACCOUNTING_POLICY <-[Discloses]- ORG <-[Discloses]- ORG",
|
| 272 |
+
"sector": "Industrials",
|
| 273 |
+
"difficulty": "medium",
|
| 274 |
+
"quality_score": 48,
|
| 275 |
+
"nodes": [
|
| 276 |
+
{"role": "start", "id": "BA", "name": "BA", "type": "ORG", "year": 2024, "idf_score": null},
|
| 277 |
+
{"role": "connector", "id": "Asset_Allocation", "name": "Asset Allocation", "type": "ACCOUNTING_POLICY", "year": null, "idf_score": 4.01}
|
| 278 |
+
],
|
| 279 |
+
"hops": [
|
| 280 |
+
{"hop": 1, "relationship": "Monitors", "source_file": "BA_10k_2024.pdf", "page_id": "page_99", "chunk_id": "chunk_1", "chunk_text": "..."},
|
| 281 |
+
{"hop": 2, "relationship": "Discloses", "source_file": "EMR_10k_2024.pdf", "page_id": "page_63", "chunk_id": "chunk_5", "chunk_text": "..."},
|
| 282 |
+
{"hop": 3, "relationship": "Discloses", "source_file": "HON_10k_2024.pdf", "page_id": "page_107", "chunk_id": "chunk_3", "chunk_text": "..."}
|
| 283 |
+
]
|
| 284 |
+
}
|
| 285 |
+
```
|
| 286 |
+
|
| 287 |
+
## Dataset Creation
|
| 288 |
+
|
| 289 |
+
### Motif families
|
| 290 |
+
|
| 291 |
+
Questions are instantiated from typed KG motifs, each under three evidence regimes with explicit
|
| 292 |
+
provenance constraints:
|
| 293 |
+
|
| 294 |
+
- **2-hop bridge** (`cross_company`, `cross_year`, `intra_doc`): cross-company uses two distinct
|
| 295 |
+
organizations with matched fiscal years; inter-year uses one organization across different years;
|
| 296 |
+
intra-document uses one document but distinct pages/chunks (or distinct facts in a shared chunk).
|
| 297 |
+
- **3-hop connector chain** (`type1`): connector pairs are chosen by IDF ranking within GICS strata, which
|
| 298 |
+
suppresses generic connectors; the linking relation must share provenance with the anchor edges.
|
| 299 |
+
- **3-hop multi-branch** (`type2`): three organization anchors (or one organization across three years)
|
| 300 |
+
each linked independently to one shared connector node.
|
| 301 |
+
|
| 302 |
+
### Generation and filtering
|
| 303 |
+
|
| 304 |
+
The pipeline generates a full corpus of **36,668** QA pairs; this benchmark is the filtered subset of
|
| 305 |
+
**15,600**. Filtering preserves the topology and sector balance of the full corpus.
|
| 306 |
+
|
| 307 |
+
### Quality control
|
| 308 |
+
|
| 309 |
+
Every generated pair is scored by an LLM validator on five dimensions (0–10 each, max 50):
|
| 310 |
+
multi-chunk dependency, financial context, answer grounding, causal-connection grounding, and natural
|
| 311 |
+
analyst quality. Pairs below **45/50** are regenerated. Independent of the score, a pair is discarded if
|
| 312 |
+
it:
|
| 313 |
+
|
| 314 |
+
- mentions page or section structure;
|
| 315 |
+
- is answerable from one evidence fact without combining the required hops;
|
| 316 |
+
- needs external knowledge not present in the evidence;
|
| 317 |
+
- asserts a fabricated causal link between companies that only share a generic term.
|
| 318 |
+
|
| 319 |
+
The rubric total is released as `quality_score` on every row.
|
| 320 |
+
|
| 321 |
+

|
| 322 |
+
|
| 323 |
+
### Human (SME) validation
|
| 324 |
+
|
| 325 |
+
Two finance SMEs independently rated a fixed-seed, topology-stratified random sample of 360 QA pairs on
|
| 326 |
+
four criteria. A separate SME judge-calibration study scored model answers on a distinct 360-pair panel
|
| 327 |
+
(1,080 assessments).
|
| 328 |
+
|
| 329 |
+
## Uses
|
| 330 |
+
|
| 331 |
+
- Evaluating LLM multi-hop reasoning over financial filings
|
| 332 |
+
- Comparing evidence-delivery strategies (KG-linked chunks vs. page windows vs. retrieval)
|
| 333 |
+
- Studying how performance changes with hop count, topology, and motif
|
| 334 |
+
|
| 335 |
+
The benchmark paper evaluates six evidence protocols: KG-linked provenance (E1), page windows (E2), page
|
| 336 |
+
windows with distractors (E3), relevance-sorted descending (E4) and ascending (E5), and semantic retrieval
|
| 337 |
+
(E6).
|
| 338 |
+
|
| 339 |
+
## Known Issues
|
| 340 |
+
|
| 341 |
+
- **2-hop intra-document labels:** `document_relationship` is set to `inter_document_cross_company` for all
|
| 342 |
+
1,500 rows in the two `*_2hop_intra_doc` partitions. Use the `topology` column instead, which is derived
|
| 343 |
+
from the partition. 3-hop intra-document rows are correct.
|
| 344 |
+
- **`question_id` is not unique.** Use `id` as the row key.
|
| 345 |
+
- **`difficulty`** is an IDF heuristic (~80% `hard`), not a calibrated difficulty measure.
|
| 346 |
+
|
| 347 |
+
## Citation
|
| 348 |
+
|
| 349 |
+
If you use this dataset, please cite:
|
| 350 |
+
|
| 351 |
+
```bibtex
|
| 352 |
+
@article{arun2026finmultihopqa,
|
| 353 |
+
title={FinMultiHopQA: Financial QA Benchmark for Reasoning with Knowledge Graph Evidence},
|
| 354 |
+
author={Arun, Abhinav and Harsh, Reetu Raj and Sarmah, Bhaskarjit and Pasquali, Stefano},
|
| 355 |
+
journal={arXiv preprint},
|
| 356 |
+
year={2026}
|
| 357 |
+
}
|
| 358 |
+
```
|
| 359 |
+
|
| 360 |
+
FinMultiHopQA is built on FinReflectKG:
|
| 361 |
+
|
| 362 |
+
```bibtex
|
| 363 |
+
@article{arun2025finreflectkg,
|
| 364 |
+
title={FinReflectKG: Agentic Construction and Evaluation of Financial Knowledge Graphs},
|
| 365 |
+
author={Arun, Abhinav and Dimino, Fabrizio and Agarwal, Tejas Prakash and Sarmah, Bhaskarjit and Pasquali, Stefano},
|
| 366 |
+
journal={arXiv preprint arXiv:2508.17906},
|
| 367 |
+
year={2025},
|
| 368 |
+
url={https://arxiv.org/pdf/2508.17906}
|
| 369 |
+
}
|
| 370 |
+
```
|
assets/hop_count_doc_type_topology.png
ADDED
|
Git LFS Details
|
assets/kg_multihop_inter_year_1.png
ADDED
|
Git LFS Details
|
assets/motif_distribution.png
ADDED
|
Git LFS Details
|
assets/sector_distribution_full_vs_filtered.png
ADDED
|
Git LFS Details
|
data/qualitative_2hop_cross_company.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:1da11ac70cdfba5e52b18c1b1fcbbef12bb80428d44ad08b27da83760b715806
|
| 3 |
+
size 5914787
|
data/qualitative_2hop_cross_year.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:f810476d0e961e264b4a2e295985a9622d82d04f48c5d02ca81195f4a73f2d6e
|
| 3 |
+
size 3815595
|
data/qualitative_2hop_intra_doc.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:3f4a953000c134f777e1933f557a30d14afdc9d678edb5f951986873084da4db
|
| 3 |
+
size 2355418
|
data/qualitative_3hop_cross_company_type1.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:9b02c40dcbd127c504fe05a2e2e8c1285307eeaf923e6083334f677212d6e7aa
|
| 3 |
+
size 2972878
|
data/qualitative_3hop_cross_company_type2.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:3941322a28cfe2c716c95debf1cd3ea1cdcc433708ff440b0246f264421586dc
|
| 3 |
+
size 9009154
|
data/qualitative_3hop_cross_year_type1.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:7c01de14c226b3cc5a3e9fc6d9c167f009acc6e345fefcb9469d936edb6b09ba
|
| 3 |
+
size 3447847
|
data/qualitative_3hop_cross_year_type2.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:ff507fef261d40e90e488ea18f3dd5c268fe1b0d1d14382b63a3b06fb6b36dd0
|
| 3 |
+
size 3514996
|
data/qualitative_3hop_intra_doc_type1.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:6008c384b702d38a73b9900d25ba2971b62b3f7f762a4d115f7f09acaa75534e
|
| 3 |
+
size 2401928
|
data/qualitative_3hop_intra_doc_type2.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:2080ca49b871928a2ec2f9b8f6bc46fecd9dec16ef12c82e76dc003102fa3632
|
| 3 |
+
size 3310498
|
data/quantitative_2hop_cross_company.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:af79a75681064f2a1fc2d660c349aacac7e18a30a3cafdc7141ca51114fa3b40
|
| 3 |
+
size 3710218
|
data/quantitative_2hop_cross_year.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:34aa0b3f1f7bd261c288675c3f95c6d16de183f93d9516aa14fa909dcd454f67
|
| 3 |
+
size 2637070
|
data/quantitative_2hop_intra_doc.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:b815e8d3ff502adf7cb613f9f90aea1839ca89cc55996e8caf06ea1e533b1152
|
| 3 |
+
size 1429143
|
data/quantitative_3hop_cross_company_type1.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:cc5aabc5563bc09c47907134f2f31330089502112b9af3bae151c35b0bdaf837
|
| 3 |
+
size 1873016
|
data/quantitative_3hop_cross_company_type2.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:d4694e5f05cc87d9917ab87e785560436f476c0783bf6984ebb3ffb60af41b4d
|
| 3 |
+
size 6412996
|
data/quantitative_3hop_cross_year_type1.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:8d3833cbd662c75376f7bf719423a8a28b7ec976d232e523ec0367cd5c29258b
|
| 3 |
+
size 2279154
|
data/quantitative_3hop_cross_year_type2.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:55aebd6e73393bd1606d7e86fe6f7e221e9a0dbae32fc8d19dec4d7381dc82d8
|
| 3 |
+
size 2120589
|
data/quantitative_3hop_intra_doc_type1.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:3406c41c3a613466f5ae4898126f41624fc6b36630151b842c7707f035efe331
|
| 3 |
+
size 1862615
|
data/quantitative_3hop_intra_doc_type2.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:8b9f10ba65b761bb5b7e9181831f1bec1875a58624134775beeae15342ea6516
|
| 3 |
+
size 2604633
|