| --- |
| language: |
| - en |
| - hi |
| license: openrail |
| task_categories: |
| - text-classification |
| - text-generation |
| - question-answering |
| tags: |
| - legal |
| - india |
| - supreme-court |
| - law |
| size_categories: |
| - 10K<n<100K |
| --- |
| |
| # Indian Supreme Court Judgments (Structured JSON) |
|
|
| ## About Dataset |
| This dataset contains fully extracted, structured JSON data for Indian Supreme Court judgments. This is a parsed, machine-readable version of the raw PDF repository, designed specifically for Natural Language Processing (NLP), RAG (Retrieval-Augmented Generation), and legal tech machine learning applications. |
|
|
| The dataset is provided in `.jsonl` (JSON Lines) format. Each row represents a single case and contains rich metadata alongside the actual extracted text, including fields like `full_text`, `sections_cited`, and `articles_cited`. |
|
|
| ## Source & Provenance |
| The original source for these judgments is the Indian Supreme Court government website API (https://main.sci.gov.in/). |
|
|
| The raw, unparsed PDF files that this text was extracted from are available in a separate companion dataset here: |
| 👉 **[Comming Soon]** |
|
|
| ## Extraction Notes & Data Quality |
| * **Extraction Pipeline:** The text was extracted utilizing a hybrid pipeline based on PyMuPDF to convert PDF structure into clean JSON. You can find the scraping and extraction pipeline code on GitHub. |
| * **Scanned/Photo Documents:** Some of the original PDFs provided by the court API are scanned images or photos of older judgments. In cases where text extraction failed or OCR could not confidently extract the text, the `full_text` field may be empty. If you need to manually review these, their original PDF documents are still available in the raw PDF dataset linked above. |
| * **Languages:** Judgments in vernacular languages are also available in this database. The metadata contains a `language` column. |
| * **Identifiers:** There is no specific unique ID provided natively by the court API for the judgments. `diary-no` and `case-no` are not strictly unique, as there can be multiple judgments for each case or diary number. Some data doesn't have a diary number; the court API assigns a `-0` for all such judgments. |
| * **Duplicates/Metadata Anomalies:** Metadata also contains examples where the same judgment is part of multiple years. For example, diary-no `17050-2006` appears with two judgment dates while the judgment PDF link is exactly the same. Some links in the raw metadata needed to be prefixed with `jonew/` to get a correct URL. |
|
|
| ## Usage |
| Because the data is in `.jsonl` format, it is extremely memory efficient to load using the Hugging Face datasets library: |
|
|
| ```python |
| from datasets import load_dataset |
| |
| # Load the dataset |
| dataset = load_dataset("json", data_files="laww_dataset.jsonl", split="train") |
| |
| # Print the full text of the first case |
| print(dataset[0]['full_text']) |
| |
