niloycste68/IndicIPRExtractiveQA

IndicIPR-QA

A Benchmark for Extractive Question Answering on Indian IPR Laws

IndicIPR-QA is an English-language benchmark for extractive question answering over Indian Intellectual Property Rights (IPR) statutes.

It contains 4,462 question–answer pairs derived from 1,116 statutory contexts covering all seven principal Indian IPR Acts.

Every gold answer is an extractive span grounded directly in its source statutory context, making the benchmark suitable for evaluating models that must locate precise and traceable answers from legal text.

---

Research Use Notice

IndicIPR-QA is designed for research and educational purposes.
It is not a legal-advice system and should not be used as a substitute for professional legal advice or authoritative statutory interpretation.

---

✨ Key Highlights

• 4,462 extractive QA pairs
• 1,116 retained statutory contexts
• Coverage of all seven principal Indian IPR Acts
• Section-level train/validation/test separation
• GPT-4o-assisted QA generation
• Nine-step automatic quality-control pipeline
• Context-balanced candidate selection
• Single- and multi-model LLM quality diagnostics
• Post-hoc human audit of 200 test QA pairs
• Evaluation with:
  • BM25
  • BERT-base
  • LegalBERT-US
  • InCaseLawBERT
  • InLegalBERT
  • Mistral 7B
  • Llama 3.1:8B
  • Qwen2.5:7B
• Public generation prompt and reproducible pipeline

---

📊 Dataset at a Glance

Property	Value
Total QA pairs	4,462
Initial statutory contexts	1,173
Retained statutory contexts	1,116
Acts covered	7
Training samples	3,710
Validation samples	408
Test samples	344
Average context length	171.3 words
Average question length	11.9 words
Average answer length	35.6 words
Language	English
Task	Extractive Question Answering
Domain	Indian Intellectual Property Law
Generation model	GPT-4o
Human audit	200 test QA pairs

---

⚖️ Indian IPR Acts Covered

IndicIPR-QA covers the following seven principal Indian Intellectual Property Rights statutes:

1. The Patents Act, 1970
2. The Trade Marks Act, 1999
3. The Copyright Act, 1957
4. The Designs Act, 2000
5. The Geographical Indications of Goods (Registration and Protection) Act, 1999
6. The Semiconductor Integrated Circuits Layout-Design Act, 2000
7. The Protection of Plant Varieties and Farmers' Rights Act, 2001

The statutory texts were collected from official Government of India documents available through the India Code portal.

---

🎯 Why IndicIPR-QA?

Legal Question Answering over statutes is different from ordinary passage QA.

Indian IPR statutes frequently contain:

• long provisions,
• nested subsections,
• definitions,
• provisos,
• exceptions,
• legal conditions,
• procedural requirements,
• rights and powers,
• penalties,
• cross-references,
• multi-clause answers.

A system must therefore do more than identify a relevant section.

It must also select the correct start and end boundaries of the answer span.

IndicIPR-QA is designed to support research on:

• exact statutory answer extraction,
• legal answer-span prediction,
• long-answer QA,
• multi-clause statutory answers,
• answer-boundary errors,
• domain-specific language models,
• legal NLP evaluation,
• grounded legal-information QA,
• generalization to unseen statutory contexts.

---

📁 Dataset Files

The benchmark files are located in the data/ directory.

File	Description	Size
merged_ip_qa.jsonl	Full curated IndicIPR-QA dataset	4,462 QA pairs
train_squad.jsonl	Training split	3,710 QA pairs
val_squad.jsonl	Validation split	408 QA pairs
test_squad.jsonl	Test split	344 QA pairs

These four files constitute the core IndicIPR-QA benchmark.

---

🧩 Dataset Splits

Split	QA Pairs
Train	3,710
Validation	408
Test	344
Total	4,462

IndicIPR-QA uses statute-section-level splitting.

The same source context does not appear in both the training and test sets.

---

📚 Per-Act Distribution

Act	Full Dataset	Test Set
Designs	173	24
Patents	1,143	69
Semiconductor	585	47
Trade Marks	1,106	88
Geographical Indications	551	51
Copyright	349	25
Plant Varieties	555	40

---

🧾 Data Format

The train, validation, and test files use a SQuAD-style extractive QA structure.

{
  "id": "cop_000001",
  "title": "Section 1: Short title, extent and commencement",
  "context": "statutory text...",
  "question": "What does sub-section (1) of Section 1 provide?",
  "answers": {
    "text": [
      "answer text"
    ],
    "answer_start": [
      0
    ]
  }
}

The answer is represented using:

• text: the gold answer span
• answer_start: the starting character position of that span in the context

---

🏗️ Dataset Construction Pipeline

IndicIPR-QA was constructed through a multi-stage pipeline:

Official Indian IPR Acts
        │
        ▼
Text Extraction & Cleaning
        │
        ▼
1,173 Statutory Contexts
        │
        ▼
GPT-4o QA Generation
        │
        ▼
5,721 Raw QA Candidates
        │
        ▼
Nine Automatic Quality Filters
        │
        ▼
4,689 Filtered QA Candidates
        │
        ▼
Context-Balanced Selection
        │
        ▼
4,462 Final QA Pairs
        │
        ▼
1,116 Retained Contexts
        │
        ▼
Section-Level Splitting
        │
        ├── Train: 3,710
        ├── Validation: 408
        └── Test: 344
        │
        ▼
LLM Diagnostics + Human Audit

---

📝 Source Corpus Preparation

The full text of the seven Indian IPR Acts was collected from official Government of India PDFs available through the India Code portal.

During preprocessing:

• table-of-contents pages were removed,
• editorial amendment markers such as Ins., Subs., and w.e.f. were cleaned,
• whitespace was normalized,
• sentence boundaries were preserved,
• statutory section boundaries were preserved where possible,
• the cleaned text was divided into paragraph-level contexts.

This process produced 1,173 initial statutory contexts.

---

🤖 GPT-4o Question–Answer Generation

GPT-4o was used to generate up to five candidate extractive QA pairs from each statutory context.

The generation procedure was designed to encourage:

• exact answer spans,
• legally specific questions,
• concise answers,
• diversity of legal question types,
• structured output,
• reduced answer leakage,
• reduced vague questioning,
• reduced near-duplicate questions.

---

Generation Settings

Setting	Value
Model	gpt-4o
Temperature	API default / not explicitly set
QA pairs requested per context	Up to 5
Initial contexts	1,173
Raw generated candidates	5,721
After automatic filtering	4,689
Final selected QA pairs	4,462
Retained contexts	1,116

System Message

You generate extractive legal QA and return only valid JSON.

---

⭐ Full GPT-4o Generation Prompt

To improve transparency and reproducibility, the complete generation prompt used for dataset construction is provided below.

The prompt is generated by build_prompt() in:

Dataset Preparation/step2_generate_qa_from_unique_contexts.py

The values inside {...} are dynamically replaced for each statutory context.

You are creating an EXTRACTIVE legal QA dataset from Indian legal text.

Generate up to {questions_per_context} high-quality extractive question-answer pairs from the context below.

STRICT RULES:

1. Every answer MUST be an exact substring of the context.

2. Every answer MUST be as short as possible while still legally correct.

3. Do NOT ask yes/no questions.

4. Do NOT ask vague questions like:
   - What is stated in this section?
   - What is laid down in this section?
   - What does this section say?

5. Prefer diverse legal question types:
   - definition
   - right
   - power / authority
   - condition
   - prohibition
   - exception
   - duration / time period
   - procedure
   - penalty / consequence

6. Avoid near-duplicate paraphrases.

7. Prefer answer lengths between 2 and 30 words.

8. Stay strictly within the given context.

9. Output JSON only.

Return JSON:

{
  "items": [
    {
      "question": "....?",
      "answer": "exact substring from context",
      "question_type": "definition|right|power|condition|prohibition|exception|duration|procedure|penalty|other"
    }
  ]
}

Act: {act}

Title: {title}

Context:
"""{context}"""

Purpose of the Prompt Constraints

The prompt was designed around several important requirements for an extractive legal QA benchmark.

Prompt Requirement	Purpose
Exact substring answers	Keep answers grounded in the statutory source
Minimal answers	Reduce unnecessarily long gold spans
No yes/no questions	Encourage answer extraction rather than classification
No vague questions	Encourage legally specific questions
Diverse question types	Increase legal-content coverage
No near-duplicates	Reduce repeated paraphrases
Stay inside context	Reduce unsupported answers
JSON-only output	Support automated processing

The generation prompt alone does not determine whether a candidate enters the final benchmark.

Every generated candidate subsequently passes through the automatic quality-control pipeline described below.

---

🔍 Nine-Step Quality-Control Pipeline

All 5,721 raw GPT-4o-generated QA candidates are processed using nine sequential automatic filters.

Filter	Criterion	Purpose
F1	Required fields	Remove incomplete records
F2	Answer in context	Ensure extractive grounding
F3	Answer length: 5–350 characters	Remove extremely short or long spans
F4	Answer covers <85% of context	Remove near-context-length answers
F5	Question quality	Enforce question form and specificity
F6	No answer leakage	Prevent the question from revealing its answer
F7	Context not truncated	Remove incomplete contexts
F8	No editorial artifacts	Remove amendment/editorial noise from answers
F9	Answer not numeric-only	Remove standalone 1–4 digit answers

The filtering stage reduces:

5,721 raw candidates
        ↓
4,689 filtered candidates

F3 and F4 account for most automatic exclusions.

After filtering, a context-balanced selection procedure reduces the candidate set to:

4,689 filtered candidates
        ↓
4,462 final QA pairs

covering 1,116 retained statutory contexts.

---

⚖️ Human Quality Audit

A post-hoc human quality audit was conducted on 200 QA pairs from the test split, covering all seven Acts.

Two law graduates independently evaluated each QA pair using three criteria:

1. Question Clarity
2. Answer Correctness
3. Answer Minimality

Each criterion used the following ordinal scale:

1 = Poor
2 = Acceptable
3 = Good

Agreement was measured using:

• unweighted Cohen's κ
• quadratic-weighted Cohen's κ

---

Human Audit Results

Criterion	Joint Acceptability	Unweighted κ	Quadratic-Weighted κ
Question Clarity	99.0%	-0.008	0.051
Answer Correctness	95.0%	0.117	0.328
Answer Minimality	75.5%	0.312	0.613

Joint acceptability requires both annotators to assign a score of at least 2.

For answer minimality:

• 75.5% were accepted by both annotators
• 82.0% received a score ≥2 from at least one annotator

Both annotators rated 36 QA pairs as non-minimal.

These pairs remain in the released benchmark because the human evaluation was conducted after dataset construction and served as a post-hoc quality audit rather than an additional filtering stage.

---

Interpreting the Human Audit

The human evaluation should be interpreted as a diagnostic quality audit, rather than exhaustive expert validation of every QA pair.

Important considerations include:

• only 200 test QA pairs were audited,
• the annotators were law graduates rather than specialist IPR practitioners,
• some long contexts were truncated in the annotation interface,
• some retained answers were correct but longer than necessary.

These limitations are reported to support transparent interpretation of the benchmark.

---

🧠 LLM-as-Judge Diagnostics

The final dataset was additionally examined using LLM-based quality diagnostics.

Single Judge

• Gemma 3:12B

Three-Model Ensemble

• Gemma 3:12B
• Llama 3.1:8B
• Mistral 7B

The judges independently assessed:

• Question Specificity
• Answer Sufficiency
• Answer Minimality
• Legal Relevance

Scores ranged from:

1 = Poor / Flagged
2 = Acceptable
3 = Good

The LLM-as-Judge results are treated as diagnostic signals and not as replacements for human evaluation.

---

🏆 Benchmark Results

All results below are reported on the 344-pair test split.

---

Lexical Baselines

System	EM	F1	Char-F1
BM25 Retrieval	0.00	20.21	77.10
BM25 Oracle Window	14.24	51.30	85.23

The large difference between BM25 and the learned QA models indicates that simple lexical matching alone is insufficient for the benchmark.

---

Fine-Tuned Encoder Models

Model	EM	F1	Char-F1
BERT-base	63.66	83.33	86.38
LegalBERT-US	65.12	84.66	86.67
InCaseLawBERT	63.08	84.37	86.40
InLegalBERT	63.66	85.02	87.26

InLegalBERT achieves the highest token-level F1 among the evaluated encoder models.

However, the differences between the encoder models are small.

A paired bootstrap significance test with 10,000 resamples and Holm correction found no statistically significant F1 difference between InLegalBERT and:

• BERT-base
• LegalBERT-US
• InCaseLawBERT

Therefore, the results should not be interpreted as strong evidence that domain-specific pretraining alone causes better performance.

---

Prompted Open LLMs

Model	Setting	EM	F1	Char-F1
Mistral 7B	Zero-shot	18.31	67.11	75.07
Mistral 7B	3-shot	24.71	68.83	76.01
Llama 3.1:8B	Zero-shot	33.43	74.78	79.61
Llama 3.1:8B	3-shot	33.43	77.02	79.78
Qwen2.5:7B	Zero-shot	29.94	77.59	81.04
Qwen2.5:7B	3-shot	34.30	76.39	78.60

Among the prompted LLMs:

• Qwen2.5:7B performs best in the zero-shot setting with 77.59 F1
• Llama 3.1:8B performs best in the 3-shot setting with 77.02 F1

The best prompted result remains 7.43 F1 points below fine-tuned InLegalBERT, which achieves the highest F1 among the evaluated models.

---

📈 Per-Act Performance

InLegalBERT performance varies substantially across the seven Acts.

Act	Test Samples	F1
Designs	24	100.00
Patents	69	98.40
Semiconductor	47	91.53
Trade Marks	88	83.83
Geographical Indications	51	83.66
Copyright	25	79.41
Plant Varieties	40	53.20

The Designs result should be interpreted cautiously because its test subset contains only 24 QA pairs.

The variation across Acts shows that answer-extraction difficulty is not uniform across the benchmark.

---

🔬 Error Analysis

Among InLegalBERT's 125 non-exact-match predictions:

Error Type	Count	Percentage
Span-boundary errors	100	80.0%
Low overlap	20	16.0%
Minor-word differences	2	1.6%
Partial overlap	2	1.6%
Completely wrong	1	0.8%

The dominant failure mode is therefore answer-boundary selection.

For example, a model may identify the correct statutory clause but extend its prediction into the next enumerated clause.

This is particularly relevant for long and multi-clause legal provisions.

---

📏 Answer-Structure Analysis

IndicIPR-QA contains relatively long answer spans.

Average Answer Length

35.6 words

Using the median answer length of 33 words:

Answer Group	InLegalBERT F1
Short answers	86.24
Long answers	83.78

Multi-clause answers also produce lower boundary-sensitive Exact Match:

Structure	EM
Single-clause answers	66.03
Multi-clause answers	56.10

Answer position also has a smaller effect:

Position	F1
Earlier answers	85.83
Later answers	84.21

These results indicate that answer length, clause structure, and answer position can influence extraction performance.

---

🌐 Generalization and Split Integrity

IndicIPR-QA uses section-level dataset splitting.

The train–test source-context Jaccard similarity is:

0.00

for all seven Acts.

Therefore, the same statutory source context does not appear in both the training and test splits.

The test set also contains unseen vocabulary.

• 11.1% of unique test-answer tokens do not occur in training answers
• 15.3% of test-question tokens do not occur in training questions
• 13.7% of test-context tokens do not occur in training contexts

The benchmark therefore includes unseen statutory contexts and vocabulary.

---

🔁 GPT-4o Question-Pattern Analysis

Source-context separation does not completely eliminate generation-style overlap.

Among the 4,462 questions:

• 4,118 are unique after lowercasing, replacing non-word-character sequences with spaces, and trimming
• 1,157 begin with what does
• 960 begin with under section
• 875 begin with what is

These recurring GPT-4o question structures may appear across both training and test splits.

Therefore, IndicIPR-QA measures model performance on:

unseen statutory contexts under the same question-generation process

and should not be interpreted as an evaluation on independently written human legal questions.

---

⚠️ Limitations

IndicIPR-QA has several important limitations.

LLM-Generated Questions

Questions were generated using GPT-4o.

Recurring linguistic structures and templates remain in the benchmark.

No Human-Written Comparison Set

There is currently no independent fully human-authored question subset for comparison.

English Only

The current benchmark covers English statutory text only.

Domain Scope

The benchmark is limited to the seven principal Indian IPR Acts.

It does not directly represent:

• other Indian legal domains,
• court judgments,
• other jurisdictions.

Extractive and Answerable QA Only

IndicIPR-QA currently focuses on answerable, single-context extractive questions.

It does not evaluate:

• unanswerable questions,
• scenario-based legal reasoning,
• reasoning across multiple sections,
• long-form generative legal QA.

Human Audit Scope

The human audit covers 200 test QA pairs.

The annotators were two law graduates rather than specialist IPR practitioners.

Truncated Audit Contexts

The human-audit interface displayed truncated versions of some long contexts.

The resulting human-quality statistics should therefore be interpreted as diagnostic rather than exhaustive estimates.

Answer Minimality

The human audit reports 75.5% joint acceptability for answer minimality.

Some retained gold answers therefore remain longer than necessary.

---

🚀 Quick Start

Install the Hugging Face datasets library:

pip install datasets

---

Load Train / Validation / Test Splits

from datasets import load_dataset

dataset = load_dataset(
    "niloycste68/IndicIPRExtractiveQA",
    data_files={
        "train": "data/train_squad.jsonl",
        "validation": "data/val_squad.jsonl",
        "test": "data/test_squad.jsonl"
    }
)

print(dataset)

---

Access a QA Example

sample = dataset["train"][0]

print("Question:")
print(sample["question"])

print("\nAnswer:")
print(sample["answers"]["text"][0])

print("\nContext:")
print(sample["context"])

---

Load the Full Dataset

from datasets import load_dataset

full_dataset = load_dataset(
    "niloycste68/IndicIPRExtractiveQA",
    data_files="data/merged_ip_qa.jsonl",
    split="train"
)

print(full_dataset)

---

📌 Example Data Point

Example from the Copyright Act:

Field	Example
ID	cop_000001
Title	Section 1: Short title, extent and commencement
Question	What does sub-section (1) of Section 1 provide?
Answer	This Act may be called the Copyright Act, 1957.

Example:

{
  "id": "cop_000001",
  "title": "Section 1: Short title, extent and commencement",
  "context": "(1) This Act may be called the Copyright Act, 1957...",
  "question": "What does sub-section (1) of Section 1 provide?",
  "answers": {
    "text": [
      "This Act may be called the Copyright Act, 1957."
    ],
    "answer_start": [
      4
    ]
  }
}

---

🧪 Reproducibility

The accompanying public repository contains the implementation used for:

• statutory-text extraction,
• preprocessing,
• GPT-4o QA generation,
• automatic quality filtering,
• context-balanced selection,
• dataset splitting,
• LLM-as-Judge evaluation,
• model fine-tuning,
• sliding-window inference,
• benchmark evaluation,
• statistical significance testing,
• error analysis,
• generalizability analysis.

Dataset

https://huggingface.co/datasets/niloycste68/IndicIPRExtractiveQA

Code Repository

https://github.com/niloycste/Indic-IPR-Dataset-

---

✅ Recommended Benchmark Protocol

For direct comparison with the reported benchmark results:

• use the official train/validation/test files,
• do not randomly resplit the dataset,
• preserve the section-level split,
• report results on the 344-pair test set,
• clearly specify evaluation metrics,
• clearly distinguish retrained models from evaluation-only changes,
• report any changes to the decoding or inference procedure.

The benchmark paper reports:

• Exact Match (EM)
• Token-level F1
• Character-level F1

---

🎓 Intended Uses

IndicIPR-QA is intended for research on:

• Legal NLP
• Extractive Question Answering
• Statutory QA
• Legal answer-span extraction
• Legal-information retrieval
• Domain-specific language models
• Long-answer QA
• Span-boundary prediction
• Error analysis
• Benchmark evaluation
• Grounded legal-information systems

---

🚫 Out-of-Scope Uses

IndicIPR-QA should not be used as:

• professional legal advice,
• a substitute for qualified legal counsel,
• an authoritative interpretation of Indian law,
• an automated legal decision-making system,
• a source for high-stakes legal decisions.

Users requiring authoritative legal information should consult the official statutory texts and qualified legal professionals.

---

📖 Citation

If you use IndicIPR-QA in academic work, please cite:

@inproceedings{islam2026indiciprqa,
  title     = {{IndicIPR-QA}: A Benchmark for Extractive Question Answering
               on {Indian} {IPR} Laws},
  author    = {Islam, M. Mohaiminul and Abulaish, Muhammad},
  booktitle = {2026 13th International Conference on Data Science and
               Advanced Analytics (DSAA)},
  address   = {New Delhi, India},
  year      = {2026}
}

The citation can be updated with the final IEEE DOI, page numbers, and publication metadata once they become available.

---

👥 Dataset Creators

M. Mohaiminul Islam

Department of Computer Science and Engineering
South Asian University
New Delhi, India

Muhammad Abulaish

Professor
Department of Computer Science and Engineering
South Asian University
New Delhi, India

---

🙏 Acknowledgment

If you use IndicIPR-QA, its benchmark results, or the accompanying implementation in your research, please cite the IndicIPR-QA benchmark paper and link to this Hugging Face repository.

---

⚠️ Disclaimer

IndicIPR-QA is released for research and educational purposes.

The benchmark contains statutory legal text and machine-generated question–answer pairs.

Although the dataset underwent:

• automatic quality filtering,
• LLM-based diagnostic evaluation,
• human quality auditing,
• model-based evaluation,

errors, ambiguous examples, repeated question structures, and unnecessarily long answer spans may remain.

The dataset and models trained on it should not be relied upon for legal advice, legal interpretation, or automated legal decision-making.

For authoritative legal information, users should consult the official statutory texts and qualified legal professionals.