Large, fully public corpora of Turkish court decisions are rare; this dataset supplies 11M+ full-text decisions that let researchers and engineers train, adapt, and evaluate legal NLP systems in Turkish without licensing friction. The collection emphasizes scale and provenance (Yargıtay, Danıştay, AYM, UYAP Emsal) rather than curated annotation, making it a foundation corpus rather than a labeled benchmark.
What Sets It Apart
- Coverage and scale: 11,045,085 unique decisions spanning 1962–2026 and ~31.5 billion characters, with the largest share from Yargıtay (≈9.82M). So what: you can pretrain or domain-adapt LLMs on authentic Turkish legal prose at scale, and build retrieval indexes that reflect real-world judicial language.
- Source-aware extraction: records include source, court, case numbers, ISO decision dates, text length and masked identifier counts. So what: enables time-sliced experiments, source-specific evaluation, and selective filtering (e.g., by court or year) without re-parsing raw HTML.
- Open licensing: released under CC0-1.0 and collected from public court search systems. So what: removes license barriers for research, commercial prototyping, and dataset redistribution.
Who It's For and Trade-offs
- Great fit if you need a large, raw legal text corpus for domain-adaptive pretraining, building Turkish legal retrieval/RAG systems, training summarizers or classifiers, or evaluating TL-specific NLP pipelines. The dataset is especially useful when scale and provenance matter more than gold labels.
- Look elsewhere if you need high-quality, hand-labeled legal annotations (NER, rationale labels, verdict-ground-truth) or a balanced benchmark for supervised tasks out of the box. Trade-offs include extraction noise from HTML→text conversion, temporal and source imbalance (Yargıtay dominates), and that short decisions were preserved so length distribution is wide.
Practical notes and cautions
- The publisher recommends streaming loads for the full corpus; Parquet shards allow targeted reads. The dataset masks common identifiers (e.g., [TCKN], [IBAN]) but does not guarantee removal of all sensitive content; verify against source systems before production use. Always treat outputs as research-grade and confirm legal conclusions against official records.
