Datasetpublic

Mkd-Yonas/keural-SFT-rebuilt

Keural SFT — Rebuilt (indentation-fixed) Rebuild of the keural-SFT mixed EN/KO instruction dataset, reprocessed from the original public sources with a fixed cleaning stage. ⚠️ Why this rebuild exists: the original mkd-chanwoo/keural-SFT was produced with a cleaner rule that collapsed consecutive spaces, which flattened all code-block indentation to one space — code samples were syntactically broken (root cause of a HumanEval regression). The corruption is not recoverable from… See the full description on the dataset page: https://huggingface.co/datasets/Mkd-Yonas/keural-SFT-rebuilt.

sourceHugging Faceotherupdated 2mo agoView on Hugging Face

0likes137downloads

Dataset Card

Keural SFT — Rebuilt (indentation-fixed)

Rebuild of the keural-SFT mixed EN/KO instruction dataset, reprocessed from the original public sources with a fixed cleaning stage.

⚠️ Why this rebuild exists: the original mkd-chanwoo/keural-SFT was produced with a cleaner rule that collapsed consecutive spaces, which flattened all code-block indentation to one space — code samples were syntactically broken (root cause of a HumanEval regression). The corruption is not recoverable from the published data, so all sources were re-collected and reprocessed. This dataset supersedes `mkd-chanwoo/keural-SFT` — do not train on the original if code ability matters.

—Total samples: 1,077,050 (11 shards; per-shard counts and sha256 in manifest.json)
—Languages: English (~70%) and Korean (~30%)
—Format: JSONL, fully-rendered ChatML in text (same schema as the original)

Row schema

json

{
  "text": "<|im_start|>user\n...<|im_end|>\n<|im_start|>assistant\n...<|im_end|>",
  "n_tokens": 342,
  "source_name": "magicoder",
  "license": "mit"
}

<|im_start|> / <|im_end|> must be registered as special tokens in your tokenizer (tokenizer.add_special_tokens({"additional_special_tokens": ["<|im_start|>", "<|im_end|>"]})).

Sources (11 of the original 14)

source_name	Origin	Final rows	License	Task/Lang
openorca	Open-Orca/OpenOrca (first 400K)	264,713	MIT	general/en
ultrachat	HuggingFaceH4/ultrachat_200k	214,329	MIT	general/en
orcamathkorean	kuotient/orca-math-word-problems-193k-korean	185,541	MIT	math/ko
mathinstruct	TIGER-Lab/MathInstruct	159,819	MIT	math/en
magicoder	ise-uiuc/Magicoder-OSS-Instruct-75K	73,119	MIT	code/en
alpaca	tatsu-lab/alpaca	46,852	CC-BY-4.0	general/en
koinstruct_qa	GSON-backup/KoInstruct-QA	45,253	Apache-2.0 †	general/ko
koinstruct_base	GSON-backup/KoInstruct-Base	42,191	Apache-2.0 †	general/ko
koalpaca	beomi/KoAlpaca-v1.1a	21,093	CC-BY-SA-4.0	general/ko
gsm8k	openai/gsm8k (main + socratic ‡)	12,057	MIT	math/en
competition_math	qwedsacf/competition_math	12,083	MIT	math/en

† The upstream GSON-backup/KoInstruct-* repositories carry no license tag (checked 2026-07-10). The apache-2.0 label here reflects MKD-internal confirmation of the original KoInstruct license; it replaces the license: unknown tags in the original keural-SFT.

‡ The collector ingests all files in the repo snapshot, so gsm8k includes both the main and socratic answer variants (same behavior as the original keural-SFT build).

Omitted vs. the original 14 sources: the three AI Hub datasets (aihub_multisession_sci 71674, aihub_multisession_social 71675, aihub_expert_qa 71304, ~139K rows) — pending AI Hub credentials and commercial-use confirmation. Note that AI Hub-tagged rows are excluded by the downstream training filter anyway.

Processing

Same 10-stage pipeline as the original (structure → clean → quality ≥ 0.6 → safety → per-source exact+MinHash dedup → ChatML → tokenize ≤ 8,192), with the cleaning stage fixed to preserve indentation (only unicode-abnormal whitespace and control characters are normalized). Verified post-build: 0 one-space-collapsed code blocks in sampled magicoder/openorca rows.

Integrity

bash

sha256sum *_shard_*.jsonl   # compare against manifest.json