language:

• en
  license: apache-2.0
  size_categories: 10K<n<100K
  source_datasets:
• czyssrs/FinQA
• NExTplusplus/TAT-QA
• PatronusAI/financebench
• takala/financial_phrasebank
• mteb/fiqa
• kensho/DocFinQA
• datumo/FinRED
  task_categories:
• text-generation
• question-answering
• text-classification
  task_ids:
• sentiment-classification
• document-question-answering
• closed-domain-qa
  pretty_name: FinGemma Unified Financial Dataset
  tags:
• finance
• financial
• reasoning
• nlp
• fingemma
• instruction-finetuning

FinGemma Unified Financial Dataset

The FinGemma Dataset is a production-quality, aggregated, and fully standardized instruction fine-tuning dataset designed primarily for training and evaluating financial large language models (such as FT_Gemma) with advanced document-grounded question answering (QA) and financial reasoning capabilities. It consolidates 8 premium financial datasets spanning arithmetic calculations, tabular comprehension, red-teaming safety, and sentiment analysis.

1. Overview & Dataset Description

This unified dataset provides 63,820 standardized question-answering, multi-turn dialogue, safety, and sentiment examples. Every row is normalized into a standard, deterministic layout containing:

• Instruction: Direct question, prompt, or safety query.
• Input: Accompanying structured context, including preceding texts, aligned markdown-rendered tables, and document metadata.
• Output: Validated ground truth answer (if available; unlabeled private test sets are gracefully mapped with blank strings).
• Metadata: Highly detailed trace metrics (formulas, original programs, sectors, filenames) as well as the complete, preserved original raw record to prevent any data loss during preprocessing.

2. Dataset Statistics

Consolidated Master Splits

These represent the merged, ready-to-train datasets consolidated from all 8 raw sources:

Aggregated Source Subsets

3. Source Datasets & Taxonomy

• FinQA: FinQA numerical reasoning dataset
• ConvFinQA: ConvFinQA conversational finance reasoning dataset
• TAT-QA: TAT-QA hybrid tabular and textual QA dataset
• FinanceBench: FinanceBench open-source financial QA sample
• FinancialPhraseBank: Financial PhraseBank sentiment dataset
• FiQA: FiQA 2018 opinion QA challenge dataset
• DocFinQA: DocFinQA financial document QA dataset
• FinRED: FinRED financial red-teaming benchmark

4. Unified Schema (v1.0)

Each record in the .jsonl split files strictly adheres to the following unified, flat JSON schema:

{
  "schema_version": "1.0",
  "dataset": "FinQA",
  "split": "train",
  "id": "ADI/2009/page_49.pdf-1",
  "instruction": "what is the interest expense in 2009?",
  "input": "Preceding text:\n...\n\nTable:\n...\n\nFollowing text:\n...",
  "output": "380",
  "metadata": {
    "filename": "ADI/2009/page_49.pdf",
    "program": "divide(100, 100), divide(3.8, #0)",
    "raw": { ... }
  }
}

5. Repository Directory Structure

The repository contains both consolidated master split files and independent subfolders for granular Loading/Filtering:

datasets/processed/
├── train.jsonl             # Consolidated master training splits
├── validation.jsonl        # Consolidated master validation splits
├── test.jsonl              # Consolidated master test splits
├── FinQA/                  # FinQA subset splits
├── ConvFinQA/              # ConvFinQA subset splits
├── TAT-QA/                 # TAT-QA subset splits
├── FinanceBench/           # FinanceBench subset splits
├── FinancialPhraseBank/    # FinancialPhraseBank subset splits
├── FiQA/                   # FiQA subset splits
├── DocFinQA/               # DocFinQA subset splits
└── FinRED/                 # FinRED subset splits

6. Intended Use

This dataset is designed specifically for:

• Fine-tuning Large Language Models on complex, multi-modal financial math and text.
• Evaluating mathematical, conversational, safety, and document understanding capacities.
• Aligning models to perform instruction-following, document comprehension, and complex financial calculations.

7. Limitations & Edge Cases

• Unlabeled Private Tests: Official private test sets for FinQA, ConvFinQA, and TAT-QA intentionally lack ground-truth answers. These are preserved with "has_ground_truth": false to support inference-level benchmarking.
• Large Context Sizes: DocFinQA contains long text Context records (often exceeding 330 KB). Attention sliding windows or robust retrieval-based chunking is recommended during training.

8. License & Terms

This merged collection is published under the Apache 2.0 License where wrapping schemas and code are concerned. Please refer to the licenses of the original datasets (e.g. Creative Commons, non-commercial licenses) for their respective data inputs.

9. Generation Info

• Dataset Schema Version: 1.0
• Last Generated: 2026-07-09 08:10:59 UTC