kaushik-harsh-99/Indian-legal-data-v3

Indian Legal Dataset V3

Overview

Indian Legal Dataset V3 is a large-scale instruction-tuning dataset focused on Indian law, constitutional law, criminal law, legal reasoning, legal drafting, and real-world legal assistance.

Compared to V2, this version expands the dataset with:

• —legal drafting instruction pairs,
• —hypothetical legal scenarios,
• —detailed IPC-focused data,
• —practical real-world legal instructions,
• —concise legal QA pairs.

After integrating the new data sources, the dataset was cleaned, normalized, and deduplicated again to reduce redundancy while preserving legal diversity.

The final dataset contains approximately 194,000 instruction-response pairs.

What's New in V3

Legal Drafting Data

V3 introduces drafting-oriented legal instruction pairs such as:

• —legal notices,
• —complaints,
• —affidavits,
• —applications,
• —petitions.

Example

{
  "instruction": "Draft a legal notice for non-payment of rent.",
  "output": "Subject: Legal Notice for Non-Payment of Rent...\n\nThis notice is issued to inform you..."
}

Hypothetical Legal Scenarios

The dataset now includes hypothetical and real-world legal situations involving:

• —property disputes,
• —criminal allegations,
• —constitutional conflicts,
• —procedural issues,
• —consumer rights.

Example

{
  "instruction": "A landlord forcefully evicts a tenant without notice. What legal remedies are available under Indian law?",
  "output": "The tenant may approach the appropriate civil court seeking restoration of possession, injunction relief, and damages..."
}

Expanded IPC Coverage

V3 significantly expands IPC-related instruction pairs including:

• —section explanations,
• —punishments,
• —offense classifications,
• —criminal law procedures.

Example

{
  "instruction": "Explain Section 420 IPC.",
  "output": "Section 420 IPC deals with cheating and dishonestly inducing delivery of property. It is punishable with imprisonment which may extend to seven years along with fine."
}

Real-World Legal Instructions

The dataset includes practical legal interactions such as:

• —procedural guidance,
• —legal rights questions,
• —law applicability scenarios,
• —everyday legal issue explanations.

Example

{
  "instruction": "What should I do if my employer refuses to pay my salary?",
  "output": "You may issue a formal demand notice, approach the labour commissioner, or file a claim before the appropriate labour authority depending on the nature of employment."
}

Legal Summarization

The dataset contains summarization-oriented instruction pairs based on:

• —court judgments,
• —legal orders,
• —judicial reasoning,
• —procedural developments.

Example

{
  "instruction": "Summarize the following legal judgment:\n\n<judgment text>",
  "output": "The Court held that demolition proceedings must comply with principles of natural justice and directed authorities to ensure rehabilitation measures before eviction."
}

Dataset Structure

Each sample follows a simple instruction-tuning format:

{
  "instruction": "...",
  "output": "..."
}

Dataset Statistics

Domains Covered

• —Constitutional Law
• —IPC (Indian Penal Code)
• —Criminal Law
• —Civil Law
• —Procedural Law
• —Legal Drafting
• —Legal Summarization
• —Legal Question Answering
• —Judicial Reasoning

Data Processing

The dataset underwent multiple preprocessing stages including:

• —JSON normalization,
• —exact deduplication,
• —system prompt removal,
• —instruction-output restructuring,
• —formatting cleanup,
• —malformed sample filtering,
• —long-context filtering.

Semantic deduplication was intentionally avoided to preserve legal diversity and nuanced legal coverage.

Differences Between V2 and V3

Intended Use

This dataset is intended for:

• —supervised fine-tuning (SFT),
• —legal-domain adaptation,
• —legal assistant development,
• —legal reasoning research,
• —legal drafting assistance.

Compatible with models such as:

• —Qwen
• —Llama
• —Mistral
• —Gemma
• —DeepSeek

Limitations

• —Some samples may still contain OCR artifacts or formatting inconsistencies.
• —Outputs should not be interpreted as professional legal advice.
• —Legal interpretations may vary depending on jurisdiction and context.

--