evolawyer/inlegalbert-sc-ner-silver

indian-sc-judgments-ner-silver

33,000 Indian Supreme Court judgments annotated with silver NER labels.

Covers 1950–2024. Each document contains the full cleaned judgment text and all automatically generated entity spans with verified character offsets. Designed as the training corpus for evolawyer/inlegalbert-sc-ner-silver.

Usage

Load with the HuggingFace datasets library:

from datasets import load_dataset

ds = load_dataset("evolawyer/indian-sc-judgments-ner-silver")
doc = ds["train"][0]

text = doc["text"]
for entity in doc["entities"]:
    span = text[entity["start_char"]:entity["end_char"]]
    assert span == entity["text"]          # always holds
    print(f"{entity['label']:<20} {span}")

Or load a specific year directly from the JSONL files:

import json
from pathlib import Path

def iter_year(data_root: str, year: int):
    path = Path(data_root) / f"{year}.jsonl"
    with open(path, encoding="utf-8") as f:
        for line in f:
            yield json.loads(line)

for doc in iter_year("data", 2019):
    print(doc["document_id"], len(doc["entities"]), "entities")

Coverage

Property	Value
Documents	~33,000 (25 test-set files withheld — see below)
Years	1950–2024
Source	Supreme Court of India
Language	English
Annotation	Silver (automatic)
Entity types	13

Entity Types

Label	Description	Primary Source
STATUTE	Indian Act, Code, or Constitutional document	Regex + gazetteer
PROVISION	Section, article, rule, or schedule	Regex
CASE_CITATION	Reporter citation for a prior case	Regex
JUDGE	Presiding judge name	Header projection
PETITIONER	Appellant / petitioner party	Header projection
RESPONDENT	Respondent party	Header projection
LAWYER	Appearing counsel	OpenNyAI NER
COURT	Court name mentioned in the judgment	OpenNyAI NER
ORG	Organisation other than a court	OpenNyAI NER
GPE	Geopolitical entity	OpenNyAI NER
DATE	Date mentioned in the judgment	OpenNyAI NER
OTHER_PERSON	Named person not in any other category	OpenNyAI NER
WITNESS	Witness name	OpenNyAI NER

Document Schema

Each file is a self-contained JSON document with the full judgment text embedded:

{
  "document_id": "2019_10_961_995_EN",
  "year": 2019,
  "text": "REPORTABLE\n\nIN THE SUPREME COURT OF INDIA\n\nCIVIL APPELLATE JURISDICTION\n...",
  "total_entities": 118,
  "entities": [
    {
      "text": "(2017) 15 SCC 720",
      "label": "CASE_CITATION",
      "start_char": 4821,
      "end_char": 4838,
      "source": "regex",
      "metadata": {}
    },
    {
      "text": "Consumer Protection Act, 1986",
      "label": "STATUTE",
      "start_char": 1204,
      "end_char": 1233,
      "source": "gazetteer",
      "metadata": {
        "canonical_name": "Consumer Protection Act, 1986",
        "act_id": "A1986-68"
      }
    }
  ]
}

text is the full cleaned judgment text (OCR-cleaned). All start_char / end_char offsets are character positions into this field.

end_char is exclusive (Python slice convention): doc["text"][entity["start_char"]:entity["end_char"]] == entity["text"] holds for every entity in every document.

source field values: regex · ner · header · gazetteer

Withheld Test Set

The 25 files listed in test_set_ids.json are excluded from this dataset, stratified across eras:

Era	Files
1950–1969	5
1970–1989	5
1990–2009	8
2010–2024	7

These files are reserved for final evaluation of the gold-annotated v1.0 model and have never been seen during training or validation.

Annotation Pipeline

Silver labels were produced by four automatic annotation stages merged into a single authoritative span set per document:

1. Metadata + header projection — case metadata JSONs (party names, judge names) mapped to character offsets via RapidFuzz fuzzy match for parties and strict regex for judges; confined to the document header zone (<1,500 chars). Produces JUDGE, PETITIONER, RESPONDENT.
2. Regex extraction — 14-pattern case citation extractor; statute and provision extractor with hierarchical linking. Produces CASE_CITATION, STATUTE, PROVISION.
3. OpenNyAI transformer NER — en_legal_ner_trf (InLegalBERT-based), run on GPU with character offset correction applied to align spans against the cleaned text. Produces LAWYER, COURT, ORG, GPE, DATE, OTHER_PERSON, WITNESS.
4. Gazetteer enrichment — full-corpus fuzzy matching against an alias dictionary built from 858 Central Act JSONs. Adds and confirms STATUTE spans missed by regex.

Merge authority rules: metadata owns the header zone for party/judge labels; regex overrides NER for overlapping citation and statute spans. All entity offsets verified: doc["text"][start_char:end_char] == entity["text"] passes for 100% of entities.

Related

• Model: evolawyer/inlegalbert-sc-ner-silver
• GitHub: evolawyer/inlegalbert-sc-ner-silver
• Website: evolawyer.com

License

Apache 2.0