Mkd-Yonas/keural-sft-mix

Keural SFT Mix (sampled)

A language/domain-balanced sample of mkd-chanwoo/keural-datasets, drawn for SFT experimentation. Each category is kept as its own config (mirroring the source layout), so the different category schemas never have to be merged.

Composition

Config	Rows	Size	Target mix
korean	1,238,411	~6.3 GB	40%
english	774,007	~2.1 GB	25%
code	619,205	~2.7 GB	20%
science	464,404	~1.5 GB	15%
Total	3,096,027	~12.7 GB	100%

How it was built

• Source: mkd-chanwoo/keural-datasets (~619M rows / ~1.3 TB), pooled across all processing runs.
• Sample size: SAMPLE_FRACTION = 0.005 (~0.5% of the source pool), allocated across categories by the target mix above.
• Method: for each category, a random subset of shards (seed = 42) was downloaded, shuffled, and truncated to the target row count — so only a small fraction of the source was ever downloaded.
• Reproducible: fixed seed = 42.

Columns

All configs share the same schema:

Column	Type
text	string
id	string
domain	string
dataset	string
timestamp	string
char_count	int64
tokens_count	int64

Usage

from datasets import load_dataset

korean = load_dataset("Mkd-Yonas/keural-sft-mix", "korean", split="train")
# also available: "english", "code", "science"

# Optionally interleave into a single SFT stream at the same ratios:
from datasets import load_dataset, interleave_datasets
cfgs = ["korean", "english", "code", "science"]
mix = interleave_datasets(
    [load_dataset("Mkd-Yonas/keural-sft-mix", c, split="train") for c in cfgs],
    probabilities=[0.40, 0.25, 0.20, 0.15],
    seed=42,
)

---

This is a sampled subset, not the full corpus. For the complete data, see the source dataset.