Viverun/jits-legal-dataset

JITS Legal Dataset

A production-ready, deterministic pipeline for processing Indian legal judgments into structured, high-quality legal datasets — with comprehensive extraction, self-citation exclusion, and multi-act statutory section detection.

---

Overview

Disclaimer: This dataset is independently created for research and engineering use. It is not an official government or judicial release and does not constitute legal advice.

The JITS Legal Dataset currently contains 3008 processed judgments in the full corpus, with 2215 rows in the current public train.jsonl release export, processed into machine-readable JSON with:

Coverage includes Supreme Court, High Court, and tribunal (CAT) decisions.

• Clean text extraction with artifact removal (Phase 1)
• Citation extraction with self-citation exclusion (Phase 2)
• Multi-act section extraction supporting 9+ statutory acts (Phase 3)
• IPC→BNS transition mapping with temporal validation (Phase 4)
• Comprehensive processing of 3008 judgments (current full-corpus build)

All outputs are reproducible, auditable, and traceable to explicit rules.

---

Intended Use

This dataset is designed to support multiple use cases:

• Research: Legal NLP benchmarking without label noise from ML
• Engineering: Structured legal data for analytics and downstream systems
• Hackathons: Ready-to-use dataset requiring no preprocessing
• Legal & GovTech: Explainable, audit-friendly legal data artifacts

This dataset is not intended to provide legal advice.

Limitations: A full-corpus deterministic audit (n=3008) verifies metadata and structural quality (court/date/case completeness, ID validity, date normalization, and referential integrity), currently reporting 69.0% metadata completeness (2077/3008), 0 non-ISO dates, 0 duplicate IDs, and 0 referential-integrity errors. This is an extraction-quality audit, not formal legal adjudication or institutional validation.

---

Dataset Metrics

Metric	Value	Notes
Full Corpus (Processed Judgments)	3008	Consolidated deterministic corpus
Release Export Rows (train.jsonl)	2215	UNKNOWN-court and unknown-year IDs excluded
Metadata Completeness	69.0% (2077/3008)	Audit metric (court + decision_date present)
Missing Court (Full Corpus)	738	Remaining UNKNOWN-court records
Missing Decision Date (Full Corpus)	384	Full-corpus missing dates
Missing Case Number (Full Corpus)	1068	Full-corpus missing case numbers
Unknown-Year IDs (Full Corpus)	165	IDs with unresolved year
Release Missing Dates	138	Missing dates in exported train.jsonl
Release Missing Case Numbers	597	Missing case numbers in exported train.jsonl
Similarity Edges	802,552	Rebuilt deterministic graph
Refined Clusters	77	Post-rebuild domain-pure clusters
Non-ISO Dates	0	Date normalization verified
Duplicate IDs	0	Uniqueness verified
Referential Integrity Errors	0	Referential checks passed

Quality Improvements

• ✅ Self-citation exclusion: No false positive citations
• ✅ Complete section extraction: Hyphenated sections (498-A, 304-B) now captured
• ✅ Section-act context: All sections linked to correct parent act
• ✅ Temporal transition guardrails: Pre-2024-07-01 judgments do not receive inferred BNS mappings
• ✅ Canonical court codes: HC/SC/TR levels with real court identifiers (ALL, DEL, BOM...)
• ✅ Tribunal separation: CAT records correctly typed as TR, not HC
• ✅ Comprehensive testing: Validated extraction modules

---

Data Generation

The dataset was generated using a deterministic pipeline that performs:

• Text normalization and stable ID generation
• Rule-based metadata and domain classification
• Statutory transition mapping for legacy cases
• Issue, citation, and landmark extraction
• Deterministic similarity graph construction

The full preprocessing, audit logic, and schemas are available in the associated GitHub repository:

👉 https://github.com/Viverun/jits-data-pipeline

Reproducibility & Provenance

Quick Verification

To reproduce the dataset generation locally:

git clone https://github.com/Viverun/jits-data-pipeline.git
cd jits-data-pipeline
pip install -e .
legal-ai pipeline && legal-ai audit --type quality

• No randomness or probabilistic models are used
• Identical inputs produce identical outputs
• Each dataset version corresponds deterministically to a specific pipeline commit

Note on Dataset Size: Earlier snapshots in this repository may show lower counts. The current deterministic corpus build contains 3008 processed judgments, while the current release export contains 2181 rows after UNKNOWN-court and unknown-year exclusion.

Changelog

v1.8 (2026-03-23)

• Added targeted CWP court fallback (Haryana marker with Himachal exclusion) when court is UNKNOWN.
• Re-ran pipeline through consolidate and regenerated train.jsonl.
• Full-corpus quality now: metadata 69.0% (2077/3008), missing court 738, missing date 384, missing case number 1068, unknown-year IDs 165, non-ISO dates 0, duplicate IDs 0.
• Release quality now: 2215 rows, missing dates 138, missing case numbers 597, unknown-year IDs excluded from export.
• Verification green: metadata tests 63/63.

v1.7 (2026-03-22)

• Added strict release gate to block unknown-year (0000) IDs in export/upload.
• Deepened decision-date fallback and year recovery in metadata/ID regeneration.
• Fixed metadata case-number regex backtracking stalls and restored stable throughput.
• Re-ran pipeline from metadata through consolidate.
• Rebuilt similarity graph (802,552 edges) and refined clusters (77).
• Referential integrity re-validated with 0 errors after graph rebuild.
• Full-corpus quality now: metadata completeness 67.4% (2027/3008), missing court 753, missing date 527, missing case number 1108, unknown-year IDs 238, non-ISO dates 0, duplicate IDs 0.
• Release quality now: 2181 rows, missing dates 154, missing case numbers 607, unknown-year IDs 0, non-ISO dates 0.

v1.6 (2026-03-22)

• Rebuilt corpus through consolidate and regenerated train.jsonl.
• Added deeper-header fallback recovery in metadata extraction.
• Added Delhi QR/order-portal detection, Orissa signed-location recovery, and Andhra Amaravati proceedings-sheet recovery.
• Added safer embedded high-court caption matching.
• Broadened case-tag support (MA, First Appeal, C.Misc., CR. WJC).
• Full-corpus quality: 66.3% metadata completeness (1995/3008), missing court 631, unknown-year IDs 321, non-ISO dates 0, duplicate IDs 0, referential integrity errors 0.
• Release quality (train.jsonl): 2377 rows, unknown-ID cases excluded 631, missing dates 382, missing case numbers 635, non-ISO dates 0.
• Verification green: metadata tests 63/63, normalize-dataset tests 3/3, pipeline-runner tests 2/2.

v1.5 (2026-03-19)

Breaking: 744 judgment IDs changed from v1.4.

• Court codes corrected across all IDs.
• Tribunal records (CAT, AFT) separated from High Court records.
• Metadata court extraction tightened; overmatched court phrases removed.
• 23 refined clusters (was 25), with 90,924 similarity edges unchanged.

v1.4 (2026-03-19)

• Regenerated all 846 records from the rebuilt raw text corpus and reran the full deterministic pipeline end-to-end.
• Fixed similarity signal extraction so statutory section overlap is populated from actual transition/section fields.
• Normalized decision dates and enforced temporal safeguards so judgments dated before 2024-07-01 do not receive inferred BNS mappings.
• Switched the pipeline to the strengthened zero-ML classifier with explicit service-domain handling and removed the prior writ-petition overclassification path.
• Wrote statutory_transitions consistently at the record top level in consolidated outputs.
• Corrected analytics compatibility for v2 extraction fields and removed the duplicate audit_landmarks method definition.
• v1.4 breaking change: 397 judgment IDs changed because domain classification was corrected. IDs containing -CV-, -CR-, or -SV- segments may differ from v1.3. If you stored v1.3 IDs externally, re-map using the judgment_id field in the new export.

Core quality metrics are computed using audit logic in: legal_ai_toolkit/analytics/audit.py::DataAuditor.audit_quality()

---

Dataset Structure

Each record contains:

• judgment_id: Stable unique identifier
• text: Full judgment text
• metadata: Court, date, case identifiers
• extractions: Structured citations, sections, issues, landmarks, and transition summaries
• classification: Rule-based domain classification
• statutory_transitions: IPC / CrPC → BNS / BNSS mappings
• provenance: Pipeline versioning and processing metadata

Schema Overview

The dataset schema is designed for clarity and ease of use:

• text (string): Raw source text of the judgment, cleaned of HTML artifacts.
• metadata (dict): Core legal metadata including court, date, and case identifiers.
• extractions (dict): Structured legal entities extracted from the text (citations, sections).
• classification (dict): Rule-based domain classification (e.g., Civil vs Criminal) with confidence scores and signal keywords.
• statutory_transitions (list): Mappings of legacy IPC/CrPC sections to new BNS/BNSS equivalents.
• provenance (dict): detailed pipeline versioning and processing timestamp for full auditability.

Note: The field annotations in the raw data is a logical grouping that acts as a container for extracted entities like citations, issues, and similarity metadata, which are detailed above under extractions and classification.

JSON Record Example

Here is a simplified example of a single record:

{
  "judgment_id": "IN-HC-ALL-2007-CR-25B3AC",
  "text": "Allahabad High Court...",
  "metadata": {
    "court": "Allahabad High Court",
    "court_level": "HC",
    "decision_date": "2007-12-11",
    "case_number": "CRIMINAL MISC. BAIL APPLICATION NO. 7936",
    "jurisdiction": "India"
  },
  "classification": {
    "domain": "criminal",
    "confidence": "high",
    "signals": {
      "criminal": ["CrPC", "IPC", "accused", "bail", "fir"]
    }
  },
  "statutory_transitions": [
    {
      "ipc": "498-A",
      "bns": null,
      "source": "pre_bns_background"
    }
  ],
  "provenance": {
    "pipeline_version": "2.0",
    "processed_date": "2026-03-19T19:35:17"
  }
}

---

Source Code

The complete data processing pipeline, schemas, and audit tools are available at:

👉 https://github.com/Viverun/jits-data-pipeline

---

Citation

If you use this dataset, please cite:

Viverun (2026). JITS Legal Dataset (v1.8). Hugging Face.

---

License

This dataset is licensed under Apache-2.0.

Community & Feedback

If you discover any noise, discrepancies, or have suggestions for improvements, please report them via the GitHub Issues page. Your feedback is crucial for refining this dataset.

Similarly, if you encounter any bugs in the processing pipeline, please contact us or open an issue on the repository. Contributions are welcome and highly appreciated!

Note by Viverun