TDK Türk Dili 2026: Held-out Turkish Perplexity Evaluation Set

A held-out text corpus for evaluating Turkish language model perplexity. The dataset was created by extracting and cleaning content from the printed 2026 issues of the TDK Türk Dili journal. It was authored by elifozgeyilmaz and last updated on June 3, 2026.

Use Cases

• Evaluate LLM perplexity based on held-out, contamination-free Turkish text.
• Benchmark Turkish language model generalization on post-training-cutoff content.
• Assess model performance on formal, literary Turkish based on the journal source.

Strengths

• Designed to be contamination-free, as the source is a printed journal not found in web-scraped pretraining corpora like CommonCrawl.
• Content is from 2026, which is likely after typical training cutoffs (around 2024).
• Source is described as 'Saf Türkçe' (Pure Turkish), suggesting a focus on formal language.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count and dataset size are unknown, which may limit suitability assessment.
• The description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

TDK Türk Dili journal, 2026 printed issues.

Collection Method

Extracted and cleaned from PDFs.

Time Range

2026

Freshness

Last updated 2026-06-03 21:22:15.