README.md

---
license: apache-2.0
task_categories:
- question-answering
- text-generation
- text-classification
- table-question-answering
language:
- en
- hi
tags:
- finance
- synthetic
- india
- Tax
- taxation
- itr
- document-ai
- financial-documents
pretty_name: Indian-Income-Tax-Returns
size_categories:
- n<1K
---
# Indian Income Tax Return Synthetic Dataset

A fully synthetic, high-fidelity dataset of **Indian Income Tax Return forms (ITR-4, ITR-5, and ITR-6)**.  Designed to support OCR, text extraction, table parsing, tax attribute detection, document intelligence, and LLM fine-tuning for structured data extraction. Each record includes a **PDF tax return** and a matching structured **JSON file** containing parsed fields.

This dataset simulates realistic taxpayer filings across:

- Individuals (with Aadhaar and father’s name)
- Partnership firms (with partner names)
- Companies (with director details)

It also models real-world financial variation including **refund cases, NIL returns, payable balances, late filings, interest under 234A/B/C, cess calculations, and filing statuses (139(1)/139(4)).**

## Dataset Details

- **Curated by:** AgamiAI Inc.
- **Language(s):** English, Hindi (romanized)
- **License:** Apache 2.0
- **Repository:** https://huggingface.co/datasets/AgamiAI/Indian-Income-Tax-Returns
- **Website:** https://www.agami.ai

**Note:** This dataset contains no real taxpayer data. All identities, values, and addresses are algorithmically generated.

## Uses

### Suitable For
- Document AI and OCR R&D
- Key-value and table extraction
- Named Entity Recognition (NER) for financial documents
- Tax computation extraction and validation workflows
- Layout-aware LLM fine-tuning (Donut, LayoutLMv3, SmolDoc, Nougat)
- Benchmarking structured extraction pipelines
- Multi-page form parsing and field alignment testing

### Not Suitable For
- Real-world tax reporting or compliance
- Fraud detection or audit-based analytics (no fraudulant patterns)
- Socioeconomic or demographic analysis
- Real-world tax policy modeling

## Dataset Structure

### Statement Formats

| ITR Form | Entity Type | Description |
|----------|-------------|-------------|
| **ITR-4** | Individuals with business income | Includes Aadhaar, father's name, basic income schedules |
| **ITR-5** | Firms / LLPs | Includes partner name, tax computation, interest schedules |
| **ITR-6** | Companies | Includes director name, business income, cess & interest |


### JSON Structure

```json
{
  "name": "NOVA EXPORTS",
  "entity": "Firm",
  "form": "ITR-5",
  "pan": "AQCPN5123F",
  "assessment_year": "2024-25",
  "filing_timestamp": "2024-11-18 14:23:55",
  "filing_type": "139(4) - Belated",
  "signatory": "Rajesh Singh",
  "dob": "14-Aug-1979",
  "address": "Sector 31, Gurgaon, Haryana - 122005",
  "income": 8235000,
  "tax": 2470500,
  "cess": 98820,
  "interest": {
    "234A": 28000,
    "234B": 51000,
    "234C": 42000
  },
  "total_payable": 2649320,
  "taxes_paid": 2650000,
  "balance": 680
}
```

## Included Variation

This dataset contains controlled variation to simulate realistic diversity found across Indian income tax filings.

- **Entity Types:** Individuals, Firms, Companies  
- **Geographies:** Major Indian metros and tier-2 cities  
- **Tax Outcomes:** Refund cases, NIL returns, and payable outstanding amounts  
- **Filing Types:** On-time, revised, and belated submissions (u/s 139(4))  
- **Document Formatting:** Variation in layout structure, addresses, dates, names, and signatory styles  

### Value Ranges

| Category | Range | Notes |
|----------|-------|-------|
| **Income** | ₹4,00,000 → ₹20,00,00,000+ | Scaled by entity type |
| **Filing Years** | AY 2022–2026 | Randomized across dataset |
| **Signatory Age** | 25–75 years | Professionally realistic |
| **Interest & Cess Calculations** | Randomized + rule-based | Includes Sections 234A/234B/234C + 4% cess |


## Dataset Creation

### Why This Dataset Exists

Real Indian income tax return datasets are highly sensitive, protected by law, and not publicly available for machine learning research.  
This synthetic dataset is created to fill that gap and enable:

- Training and benchmarking **Document AI / OCR systems**
- Layout-aware extraction and key-value pair modeling
- Building and testing fintech and compliance automation systems
- Research and education in AI for structured financial documents
- Evaluation of LLMs on financial form understanding

---

### How It Was Generated

This dataset is fully algorithmically constructed using controlled randomness and rule-based logic, including:

- **Procedural synthetic generation of taxpayer profiles**
- **Rule-driven fiscal computation and formatting**
- **Entity-specific metadata modeling** (Individuals, Firms, Companies)
- **Probabilistic variation in names, filing types, and geographies**
- **Digitally rendered PDF layouts with watermark and formatting diversity**
- **JSON output aligned directly with visible PDF values**

All data is artificial — no real taxpayer information, PAN numbers, Aadhaar values, or confidential records were used.


### What's Included

- **Account holders:** Business entities (companies, partnerships, corporations)
- **Transaction patterns:** B2B payments, employee salaries, vendor payments, business expenses
- **Regional diversity:** Major Indian metros
- **Temporal patterns:** Quarterly statements, monthly salary cycles, vendor payment patterns

## Limitations

## Limitations

1. **No real taxpayer data** — Does not reflect true economic or demographic patterns  
2. **Simplified tax logic** — Does not include deductions (80C, HRA), exemptions, MAT/AMT, surcharge tiers, or capital gains schedules  
3. **Entity coverage scope** — Includes only Individuals (business income), Firms, and Companies; excludes trusts, NGOs, foreign asset disclosures, or complex corporate filings  
4. **Format coverage** — Modeled on common ITR layouts; does not represent every version, annexure, or legacy revision  
5. **Synthetic reasoning constraints** — Does not cover all real-world edge cases  
6. **OCR realism** — May not include all real-world OCR challenges
7. **Regulatory compliance** — Not suitable for legal, taxation, or compliance workflows without real data validation  


This dataset is for structure and format learning, not behavioral modeling. Always validate on real data before production deployment.

## Citation

**BibTeX:**

```bibtex
@dataset{indian_income_tax_return_synthetic_2025,
  author = {AgamiAI Inc.},
  title = {Indian Income Tax Return Synthetic Dataset},
  year = {2025},
  publisher = {HuggingFace},
  url = {https://huggingface.co/datasets/AgamiAI/Indian-Income-Tax-Returns}
}
```

**APA:**

AgamiAI Inc. (2025). *Indian Income Tax Return Synthetic Dataset* [Data set]. HuggingFace. https://huggingface.co/datasets/AgamiAI/Indian-Income-Tax-Returns

## Glossary

## Glossary

**Indian Taxation Terms:**
- **AY (Assessment Year):**  
  The year in which income earned in the previous financial year is assessed and taxed.
- **FY (Financial Year):**  
  The period in which income is earned — from **1 April to 31 March**.
- **ITR (Income Tax Return):**  
  The official form used to report income, taxes paid, and tax liability.
- **PAN (Permanent Account Number):**  
  A 10-character alphanumeric tax identifier issued by the Government of India.
- **Aadhaar:**  
  A 12-digit resident identification number issued by UIDAI (included only for individuals).
- **ITR-4:**  
  Return form used by individuals/HUFs declaring presumptive business income.
- **ITR-5:**  
  Return form applicable for partnership firms, LLPs, and certain associations.
- **ITR-6:**  
  Return form for companies (except those claiming exemption under Section 11).
- **139(1):**  
  Return filed **on or before** the due date.
- **139(4):**  
  **Belated return** filed after the due date.
- **Health & Education Cess:**  
  A mandatory **4%** surcharge applied on calculated tax liability.
- **Self-Assessment Tax:**  
  Tax paid by the taxpayer before filing, when advance tax/TDS isn't sufficient.
- **Refund:**  
  Amount returned to the taxpayer if tax paid exceeds the computed liability.
- **Tax Payable:**  
  Outstanding tax amount due after accounting for advance tax, TDS/TCS, and credit adjustments.
- **Interest u/s 234A:**  
  Charged for **late filing** of the income tax return.
- **Interest u/s 234B:**  
  Levied for **not paying adequate advance tax**.
- **Interest u/s 234C:**  
  Applied for **late installment payment** of advance tax.
- **Verification Section:**  
  Mandatory declaration signed digitally or manually by the taxpayer or authorized signatory.


## More Information

### About AgamiAI

AgamiAI builds private AI solutions for enterprises where privacy, accuracy, and compliance are critical. Specialized in Finance, Healthcare, Legal, and Consulting.

Visit: **https://www.agami.ai**

### File Structure

Each ITR Document includes:
- `[ITR_id].pdf` - Scanned Income Tax Return (first two pages)
- `[ITR_id].json` - Structured data with full metadata

### Related Datasets

Part of AgamiAI's Indian Financial Documents collection:
- Indian Bank Statements (https://huggingface.co/datasets/AgamiAI/Indian-Bank-Statements)
- **Indian Income Tax Returns** (this dataset)
- Indian GST Documents (coming soon)
- Indian Audited Financial Documents (coming soon)

### Contact

- **Website**: https://www.agami.ai
- **HuggingFace**: https://huggingface.co/AgamiAI

---

**Version:** 1.0.0 | **License:** Apache 2.0 | **Last Updated:** December 2025

**Privacy Notice:** Entirely synthetic data. No real personal or financial information included.