SFT-General-Japanese-60K

Welcome! 👋

Need clean, natural Japanese conversations for supervised fine-tuning? You are in the right place. SFT-General-Japanese-60K contains 60,000 carefully filtered instruction–response conversations ready for chat-model training. It combines the practical breadth of open Japanese SFT data with transparent gates for safety, recency, formatting, language consistency, and redundancy—so you can focus on training rather than untangling noisy examples.

Overview

Property	Value
Records	60,000 two-turn conversations
Language	Japanese (ja)
Unit	one user instruction followed by one assistant response
Split	train
Formats	UTF-8 JSONL and Zstandard-compressed Parquet
Shards	12 JSONL shards of 5,000 rows each
Upstream	llm-jp/magpie-sft-v1.0 at main
License	Apache-2.0
Verified source SHA-256	acdb6e6baa892402581213cd0ee07284edfe7ee7d7ec6229e498ad04f649b558

Data format

Every JSONL line is a complete independent training example. The messages array follows the common Hugging Face chat format and always contains exactly one user turn and one assistant turn.

{
  "id": "sft-general-ja-00001",
  "language": "ja",
  "domain": "technology_and_programming",
  "messages": [{"role": "user", "content": "..."}, {"role": "assistant", "content": "..."}],
  "quality_scores": {"information_density": 4, "...": 4},
  "quality_checks": {"exact_deduplication": true, "near_duplicate_prompt_simhash": true},
  "provenance": {"source_dataset": "llm-jp/magpie-sft-v1.0", "source_revision": "main", "source_record_id": 0, "source_license": "Apache-2.0"}
}

data/train-00000-of-00012.jsonl through data/train-00011-of-00012.jsonl are equivalent 5,000-row shards. data/train.parquet is the equivalent single-file Parquet release.

Domain coverage

Selection uses a transparent keyword taxonomy and round-robin balancing: lower-volume domains are retained first; remaining choices rotate through the least-represented available domains. This prevents generic explanatory data from drowning out the long tail.

Domain	Records
arts_media_and_entertainment	429
business_and_career	1,302
daily_life_and_relationships	85
design_and_craft	353
education_and_learning	9,985
food_travel_and_hobbies	713
general_instruction	594
language_and_translation	12,902
mathematics_and_logic	1,460
philosophy_and_ethics	70
planning_and_productivity	1,794
reasoning_and_decision_making	383
science_and_environment	12,901
society_history_and_culture	773
technology_and_programming	12,694
writing_and_communication	3,562

Quality, safety, and anti-redundancy controls

Every released record has a score of at least 4/5 on all twelve recorded dimensions. Scores are deterministic, inspectable release signals rather than ungrounded model ratings.

Dimension	Release mean	Gate
information_density	4.82/5	≥4/5
structure_readability	4.83/5	≥4/5
length_fit	5.00/5	≥4/5
language_naturalness	4.93/5	≥4/5
safety_compliance	5.00/5	≥4/5
grammar_expression	4.00/5	≥4/5
domain_representativeness_diversity	4.99/5	≥4/5
training_signal_clarity	4.91/5	≥4/5
knowledge_value_novelty	4.00/5	≥4/5
reasoning_value	4.82/5	≥4/5
factual_reliability	4.00/5	≥4/5
instruction_value	4.91/5	≥4/5

The release rejects malformed conversations; invalid lengths; insufficient Japanese; URLs/emails; sensitive, dangerous, high-stakes, time-sensitive, or fact-heavy requests; raw chain-of-thought prompts; placeholders; non-answers; assistant identity/meta leakage; role-format leakage; prompt echoing; and repeated phrases.

To reduce redundancy, the pipeline uses normalized exact-pair deduplication and a character 3-gram SimHash screen on user instructions. A prompt with Hamming distance ≤5 from an already selected prompt inside a shared 16-bit LSH band is not retained. The full rule list and counts are in audit/quality_rubric.json and audit/rejection_summary.json.

Raw private reasoning is intentionally excluded. reasoning_value means concise, user-visible ordered explanations and decision support only.

Important limitation

factual_reliability = 4 means a record passed the risk screen for volatile, fact-heavy, and high-stakes content. It does not mean every statement received independent fact verification. Do not use this dataset as the sole basis for medical, legal, financial, safety-critical, or other high-impact systems without domain review.

Usage

Load Parquet

from datasets import load_dataset
dataset = load_dataset("parquet", data_files="data/train.parquet", split="train")

Load JSONL

from datasets import load_dataset
dataset = load_dataset("json", data_files="data/train-*.jsonl", split="train")

Apply a chat template

def format_example(example, tokenizer):
    return tokenizer.apply_chat_template(example["messages"], tokenize=False, add_generation_prompt=False)

Keep the messages roles intact, train only on the assistant turn when your SFT setup supports response masking, and respect the provenance and license fields when redistributing examples.

Data source and provenance

This release is a filtered derivative of llm-jp/magpie-sft-v1.0, a Japanese instruction/response dataset published by LLM-jp using the Magpie method. The source file was verified against the SHA-256 above before filtering. Every output row carries its source dataset, revision, upstream record ID, and source license in provenance. See ATTRIBUTION.md and audit/build_manifest.json for the full provenance chain.

License and attribution

This derivative is distributed under Apache License 2.0. Retain LICENSE, ATTRIBUTION.md, and row-level provenance when you redistribute or derive data. The license does not grant rights to use contributor trademarks.

Repository layout

Path	Description
data/	publishable JSONL shards and Parquet file
audit_report.html	local, human-readable audit report
audit/build_manifest.json	build inputs, hashes, counts, score means, and domain distribution
audit/quality_rubric.json	scoring and acceptance rules
audit/rejection_summary.json	rejection counts by rule
ATTRIBUTION.md, LICENSE	attribution and licensing material

Reproducibility

The build uses fixed seed 20260910. The report contains fixed-seed sample records; audit files contain the source hash, filtering rules, selected count, domain counts, and rejection statistics. To report a concern, include the row id and provenance.source_record_id.