TilQazyna/Til-mini-1B
Text Generation • 1.0B • Updated • 1
Datasets:
You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
Til-Corpus is released for research and development of Kazakh and multilingual language models. By requesting access you agree to: (1) use the data for lawful research / model-training purposes; (2) respect the upstream source licenses (the corpus aggregates permissively-licensed and openly-released web/code/math data); (3) not attempt to de-anonymise or re-identify individuals; (4) cite the Til program if you publish results.
Log in or Sign Up to review the conditions and access this dataset content.
Til-Corpus
A Kazakh-first multilingual pretraining corpus for the Til foundation-model program.
One repository, five loadable domains (configs): Kazakh, Russian, English, code, math.
Cleaned text, ready for tokenization and language-model pretraining. The matching tokenizer
(TilQazyna/Til-Tokenizer-128k) is
bundled at the repo root (tokenizer.json).
What this is
A single, deduplicated, sentence-cleaned corpus assembled to pretrain a ~1B-parameter Kazakh model while keeping strong Russian / English / code / math ability. Kazakh is the priority and the moat; the other domains give cross-lingual and reasoning coverage. ~71.5M documents, ~78 GB of parquet.
Configs & loading
from datasets import load_dataset
kk = load_dataset("TilQazyna/Til-Corpus", "kk", split="train", streaming=True)
code = load_dataset("TilQazyna/Til-Corpus", "code", split="train", streaming=True)
Each row: text (the document), source (upstream dataset), lang, plang (programming
language, code only).
| config | documents | content |
|---|---|---|
kk |
38,270,891 | Kazakh web, news, books, legal, PDF, parallel, wiki |
en |
17,256,542 | English educational + web |
ru |
7,464,278 | Russian web |
code |
4,450,016 | permissively-licensed source code, many languages |
math |
4,073,805 | math web text + curated math |
Sources (per domain)
- Kazakh — HPLT 2.0, mC4, CulturaX, GlotCC, multidomain-kazakh, finepdfs (PDF books / legal / education), KazParC (parallel), MADLAD-400 v1.5, Leipzig.
- Russian — FineWeb-2 (
rus_Cyrl) + C4 (ru). - English — FineWeb-Edu + C4 (en).
- Code —
github-code-clean, filtered to permissive licenses (MIT / Apache-2.0 / BSD / ISC …). - Math — OpenWebMath + FineMath.
Cleaning methodology
- Web (kk / ru / en) — sentence-level salvage: documents are split into sentences, low-quality sentences are dropped, good sentences kept (so a noisy document is repaired, not discarded — important because Kazakh text is scarce). Exact-hash deduplication.
- Code / math — whole-document (sentence-splitting would break code and formulas): exact-hash dedup on normalized text, minimum-length filter (code ≥ 80 chars, math ≥ 200), text kept verbatim.
- Parallelized per shard across cores.
Recommended training mix
Domain ratios are a train-time knob (up/down-sample). A balanced, Kazakh-aware blend:
Kazakh 15 / English 35 / Russian 18 / Code 20 / Math 12 (by token share)
Kazakh is upsampled from its unique pool; English is the largest consumed share. At a ~40B-token budget every domain stays under ~2 epochs (Chinchilla-healthy for ~1B params).
Tokenizer
tokenizer.json at the repo root is Til-Tokenizer-128k: byte-level BPE, 128K vocab, never
emits <unk>, single-digit numbers, agent/tool-call structure, and optional Kazakh morpheme
segmentation (boundary recall 0.896). Pre-tokenized token streams are not stored yet; tokenize
the text configs with this tokenizer.
Intended use & limitations
- Use — pretraining / continued-pretraining of Kazakh and multilingual LMs; tokenizer fitting; data research.
- Limitations — web text carries noise and bias despite cleaning; CC-derived sources (mC4 / CulturaX / GlotCC / C4 / MADLAD) overlap partially and are not cross-deduplicated across sources (a global fuzzy dedup is left to the training pipeline); sentence-salvage on Russian web removed a large fraction of noisy sentences. No benchmark/eval sets are included, but decontaminate against your evals before training.
License & access
Aggregates permissively-licensed and openly-released data; redistributed under MIT for the added cleaning/structure. Gated (manual approval) — request access and accept the terms. Respect the upstream source licenses for downstream use.
Provenance
Til program, experiments exp053–exp061. Companion tokenizer: TilQazyna/Til-Tokenizer-128k.
- Downloads last month
- 36