Til Corpus: A Multilingual Text Dataset for Language Model Pretraining

Til Corpus is a multilingual text dataset hosted on HuggingFace by TilQazyna. It contains a large volume of text data, with a tiered structure including raw, clean, and premium subsets. The dataset was last updated on 2026-06-20.

Use Cases

• Pretraining language models based on the large volume of multilingual text.
• Fine-tuning models for specific languages based on the dataset's Kazakh language content.
• Benchmarking text processing and cleaning pipelines based on the tiered data structure.
• Studying multilingual text distributions based on the dataset's scale.

Strengths

• The dataset uses a tiered quality system with 'premium' containing over 54 million rows.
• A sample keep rate of 54.4% suggests a quality filtering process was applied.
• The dataset is described as multilingual, with Kazakh language explicitly mentioned.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row counts are provided for tiers but the total dataset size is unknown.
• License information is unavailable, which may restrict usage.

Provenance

Source

TilQazyna on HuggingFace.

Collection Method

The description mentions a 'tiered' method using a 'sample-calibrated' approach with judged samples and cheap-signal tiering.

Freshness

Last updated 2026-06-20 00:16:44; freshness should be verified.