Datasets:

TilQazyna
/

Til-Instruct

language:
  - kk
  - ru
  - en
license: apache-2.0
task_categories:
  - text-generation
tags:
  - kazakh
  - kk
  - instruct
  - sft
  - til
  - judged
  - multi-tier
pretty_name: Til KK Instruct (judged, tiered)
configs:
  - config_name: raw
    data_files:
      - split: train
        path: data/raw/*.parquet
  - config_name: clean
    data_files:
      - split: train
        path: data/clean/*.parquet
  - config_name: premium
    data_files:
      - split: train
        path: data/premium/*.parquet

Til KK Instruct — judged & tiered

Unified Kazakh instruct SFT from TilQazyna til-kk-* + stukenov sozkz-instruct/*. Every row judged by Qwen3-VL → score (1-5), category (30-way). lang = script-detected. Columns: messages, task, source, src_lang, score, category, lang.

Tiers (configs)

raw — all 2563966 rows (any quality).
clean — score≥3: 2114016 rows.
premium — score≥4: 1581787 rows.

Score distribution

score	rows
0	5
1	428607
2	21338
3	532229
4	184264
5	1397523