Japanese RP datasets: Kyalulu Synthetic Roleplay Japanese AIO

Kyalulu’s official data collection#

Naotaka0127 is the official Hugging Face account of Kyalulu’s creator.Kyalulu Synthetic Roleplay Japanese AIO beta is a public collection for Japanese roleplay (RP), character dialogue and creative-writing research. Japanese is a focus; related English, Chinese, German and Arabic materials are also included.

The original AIO preserves third-party source files and provenance. Its SFW/NSFW derivatives partition individual conversation or text records. “Synthetic Roleplay Japanese AIO” is the project name; the archive also includes forum- and novel-derived material.

Choose among the three repositories#

Counts come from the published cards and partition reports. Sources overlap between derivatives, and files and records are different units; do not add these figures into an independent conversation count. Collection closure and partition rules are dated October 16, 2026 (Japan time). This guide’s source review date is October 17, 2026.

SFW/NSFW candidates and held records#

conservative-rules-v1 scanned 1,951,374 records from 24 canonical sources, reporting 13,382 SFW candidates, 78,928 NSFW candidates and 1,859,064 held records. Held payloads remain in the original AIO and are not copied into the derivative repositories.

Rules inspect strings throughout each original record, including all utterances, system instructions, character settings and scene settings. Malformed or empty records, possible personal information, age/consent review candidates and ambiguous content are held. A mixed or unreviewed source does not become SFW simply because no keyword matches.

Identical settings plus normalized utterances are represented once; formatted versions, redistributions and evaluation splits are not counted as new conversations. Groups with conflicting SFW/NSFW classifications are held. Partitioning does not restrict language or produce a Japanese-only corpus.

SFW denotes an automated candidate classification.It does not establish suitability for all ages, semantic classification of every record, age/consent clearance for NSFW candidates or complete human review. See the partition policy and report.

Load records with Python#

Install datasets in your chosen Python environment, then use configuration all and split train. This SFW example pins the revision reviewed for this guide.

import json
from datasets import load_dataset

ds = load_dataset(
    "Naotaka0127/Kyalulu-Synthetic-Roleplay-Japanese-SFW-AIO-beta",
    "all",
    split="train",
    revision="47c4d625fd974d1553bf492a5e0e4a7088854a59",
)
original = json.loads(ds[0]["source_record_json"])
print(ds[0]["source_repo"], ds[0]["source_file"], ds[0]["source_row"])

To select one source, replace all with its source name, changing / to --. An SFW example is Aratako--Synthetic-Japanese-Roleplay-SFW-DeepSeek-V3-0324-20k. Review sources and conditions in SOURCES.md.

For NSFW candidates, change the repository to Naotaka0127/Kyalulu-Synthetic-Roleplay-Japanese-NSFW-AIO-beta and revision to c77c5e573e2779b61dd879ca41d5f3dff65dbec5. The original AIO is a multi-format archive: use its README’s selective snapshot_download() example to retrieve the required source and provenance files.

Trace original text and provenance#

Original schemas include messages, conversations, posts and text. Decode source_record_json and separately adapt it to your trainer’s format. train is a distribution split name; it does not establish training suitability or independence from evaluation data.

Source licenses and quality evidence#

All three repositories declare other for source-specific licenses. Publication by the official account does not replace upstream conditions. Review the SFW source list, NSFW source list and upstream README/license files, retaining attribution, noncommercial and ShareAlike requirements where applicable.

The original AIO records checksums and related provenance for 367 files from all 33 sources. Initial work included structural checks of 23 sources and excerpt review of 510 samples from 17 original corpora. Ten later sources primarily received structural, count and derivation checks. The SFW/NSFW partition is a separate deterministic rule pass; improvements after training have not been established.

Before training, check languages and original schemas, select records suitable for your purpose and source conditions, and separate training/evaluation while considering shared instructions and approximate duplicates. See migration for importing character cards into Kyalulu, and CharacterBench for the separate evaluation corpus.

Snapshots reviewed for this guide#

The partition’s parent data revision is 9137cc70a4bb4a40b319b74db0f359f05be91f44 in the original AIO. Keep that distinct from the public documentation revisions reviewed here. The original AIO routing index tracks destinations and duplicate references for all scanned records.