bubbles01/legal-document-summarization

Legal Document Summarization Dataset

This dataset was prepared for a Legal Document Summarization project using Indian legal judgments.

Dataset

The dataset contains training and test records created from legal documents. Each record consists of a legal text chunk and its corresponding aligned summary.

Fields

• —doc_id — Document identifier
• —chunk_id — Chunk identifier within the document
• —section — Section associated with the chunk
• —chunk_text — Input legal text
• —ligned_target — Target summary for the chunk
• —split — Dataset split ( rain or est)
• —para_range — Paragraph range covered by the chunk
• —matched_sentences — Summary sentences aligned to the chunk
• —lignment_scores — Cross-Encoder alignment scores
• —matched_labels — Labels associated with the aligned summary sentences
• —structure_fallback — Whether a structural fallback was used
• —chunk_index — Index of the chunk
• — chunktokens_est — Estimated number of input tokens
• — targettokens_est — Estimated number of target tokens

Data Preparation

The legal documents were processed using section detection and chunking. Summary sentences were then aligned with relevant document chunks using a Cross-Encoder.

The resulting pairs are used as supervised data for training a sequence-to-sequence summarization model such as BART.

Task

The main task is legal document summarization:

Legal document chunk → Concise aligned summary

Splits

The training and test documents are kept separate during dataset construction. The dataset pipeline also performs split-isolation checks to prevent cross-split duplicate text.

Intended Use

This dataset is intended for academic experimentation and evaluation of legal document summarization models.

Repository Contents

• —rain.jsonl — Training records
• —est.jsonl — Test records

Acknowledgement

The source legal documents and reference summaries are used for academic NLP experimentation.