Quata2 Data v5 Clean is the release view of the Quata2 v5 supervised fine-tuning mix: a compact, multilingual SFT set built from teacher-written conversations, filtered for quality, and scrubbed where user-derived text was most likely. It is designed to train models, not to flatter them on a leaderboard.

Highlights

21,857 rows, seven configs — agent, chat, code, ifeval, math, serbian, uncertainty
English and BCMS — Serbian, Croatian and Bosnian included, not bolted on
Verified where it counts — code runs, maths matches, instruction constraints are checked
Scrubbed where it matters — chat and Serbian rows had emails, phones, public IPs and device IDs removed
No raw pools, no reasoning traces — only clean messages rows, source labels and categories
Drop-in configs — load any subset with load_dataset(repo, config, split="train")

Get started

from datasets import load_dataset

repo = "M1n1A1/quata2-data-v5-clean"

agent = load_dataset(repo, "agent", split="train")
chat = load_dataset(repo, "chat", split="train")
serbian = load_dataset(repo, "serbian", split="train")

print(chat[0])

Each config has a train split. There is no official test split; create your own held-out split before evaluation.

What's inside

Seven configs, one schema. The mix is deliberately uneven: uncertainty and maths carry the most rows, while Serbian stays small and seed-derived.

Config	Rows	Turns	Category	What it teaches
`agent`	2,299	multi	`tool_use`	multi-turn tool use; Wikipedia-grounded rows removed
`chat`	4,086	1-2	`chat`	WildChat prompts with teacher replies; PII redacted
`code`	3,495	1	`code`	coding tasks checked against tests
`ifeval`	1,532	1	`instruction_following`	instruction-following prompts and constraints
`math`	3,730	1	`math`	word and competition maths with checked answers
`serbian`	1,520	1	`serbian`	Serbian/BCMS seed-derived rows only
`uncertainty`	5,195	1	`honesty`	honesty and underspecification tasks
Total	21,857

agent rows contain system, user, assistant, and tool turns. All other configs are single-turn user/assistant rows, except some chat rows that contain two user/assistant turns.

The serbian config follows the Serbian/BCMS training mix; it contains Serbian, Croatian, and Bosnian text, not only Serbian.

Format

Every row is a JSONL object with a messages list, a coarse source label and a category.

Field	Type	Description
`messages`	`list[dict]`	Conversation turns. Each turn has `role` and `content`.
`source`	`string`	Coarse provenance label such as `v5_chat/English`. Not a unique ID.
`category`	`string`	One of `tool_use`, `chat`, `code`, `instruction_following`, `math`, `serbian`, `honesty`.

Example chat row:

{
  "messages": [
    {"role": "user", "content": "Explain why the sky is blue in two sentences."},
    {"role": "assistant", "content": "..."}
  ],
  "source": "v5_chat/English",
  "category": "chat"
}

Agent rows use the same schema with additional roles. A typical trajectory is:

system -> user -> assistant (tool call) -> tool -> assistant

The source field is a coarse label, not an upstream row ID. Do not use it for deduplication.

Files

.
├── agent/train.jsonl
├── chat/train.jsonl
├── code/train.jsonl
├── ifeval/train.jsonl
├── math/train.jsonl
├── serbian/train.jsonl
├── uncertainty/train.jsonl
├── LICENSE
├── NOTICE
├── manifest.json
└── README.md

manifest.json contains per-config row counts, byte sizes, and SHA-256 checksums.

Sources and licenses

The release is a filtered, teacher-rewritten derivative of the sources below. The packaging and annotations are CC-BY-4.0; upstream material remains under its own terms.

Config	Upstream material	License	How it is used
`agent`	Quata2 tool schemas and templates	CC-BY-4.0	synthetic trajectories; Wikipedia-grounded rows removed
`chat`	WildChat-1M	ODC-BY-1.0	prompts; teacher writes replies
`code`	OpenCodeInstruct, self-oss-instruct, code_contests	CC-BY-4.0 / ODC-BY-1.0 / CC-BY-4.0	prompts and tests; teacher answers must pass tests
`ifeval`	WildChat-1M	ODC-BY-1.0	prompts; teacher writes constrained replies
`math`	GSM8K train, NuminaMath-1.5	MIT / Apache-2.0	problems; teacher answers must match references
`serbian`	FineWeb-2 and the project's WormT crawl	ODC-BY-1.0 / project crawl	seed passages; teacher rewrites them into user messages
`uncertainty`	CoCoNot	ODC-BY-1.0	prompts; teacher writes replies

The teacher model is Qwen3.8-27B (Apache-2.0). Full attribution and upstream terms are in NOTICE. If you redistribute this dataset, keep NOTICE with it.

GSM8K: only the train split was used for training data. The test split was used only for 13-gram decontamination. Evaluate on the official test split, never on these train rows.

Serbian seed passages include public web text collected by the project's WormT crawler. Upstream pages remain subject to their own terms.

Quality gates

Rows only made the cut when they cleared the checks relevant to their config.

Teacher replies were filtered by a rubric judge; weak rows were dropped.
Code answers were accepted only when they passed the task's tests.
Maths answers were accepted only when the final answer matched the reference.
Instruction-following rows were checked against their programmatic constraints.
Serbian/BCMS rows were checked for dialect and script consistency.
Prompts were decontaminated with 13-gram overlap checks against GSM8K test, IFEval, HumanEval/MBPP/LiveCodeBench, and the project scorecard.
Rejected rows, raw candidate pools, teacher reasoning traces, and source IDs are not included.

The cleanup, measured

Check	Result
Rows kept	21,857
Schema/role/content errors	0
Residual chat/Serbian emails, phones, public IPs, device IDs	0
Wikipedia-grounded `agent` rows removed	96
Credential-giveaway `chat` rows removed	3
Translated/non-`srs_` Serbian rows removed	1,164
Email addresses redacted in chat/Serbian	19
Phone numbers redacted in chat/Serbian	6
Public IPv4 addresses redacted in chat/Serbian	6
Device identifiers redacted in chat	2

The remaining agent rows may mention wikipedia in their tool schema, but they contain no Wikipedia article content or URLs.

Privacy

Best-effort PII redaction was applied to chat and serbian, the configs most likely to contain user-derived text. Placeholders are [EMAIL_REDACTED], [PHONE_REDACTED], [IP_REDACTED], and [DEVICE_ID_REDACTED].

Private and documentation IP ranges, public DNS resolvers, and version-like strings were left intact. Synthetic/example emails in code and ifeval, and public hotline numbers in uncertainty, were not redacted because they are not user contact data.

This is not a formal anonymisation or safety review. Names, indirect identifiers, or overlooked PII may remain. If you find a privacy issue, please open an issue or contact the maintainer.

Use it for

Continued pretraining or supervised fine-tuning of instruction-following language models.
Research on tool use, instruction following, honesty, and BCMS language data.
Building evaluation sets for similar tasks, provided you decontaminate against this release.

Don't use it for

Ground truth, factual reference, or safety-critical decisions.
Medical, legal, or financial advice.
Redistribution of upstream text outside the upstream licenses.
Benchmark evaluation on train rows, including GSM8K train.

Honest caveats

This is a teacher-generated SFT set, not a verified knowledge base.

Teacher-generated answers can be incorrect, hallucinated, biased, or stylistically inconsistent.
The data is not exhaustively safety-filtered; offensive or unsafe text may remain.
Serbian seed material includes public web text from multiple sources; downstream use should respect upstream page terms.
PII scrubbing is regex-based and incomplete.
There is no official test split, and the data may overlap public benchmarks despite decontamination.

Details

Format: JSONL, one conversation per line, train split only
Teacher: Qwen3.8-27B (Apache-2.0), served locally for generation
Dataset license: CC-BY-4.0 for packaging and annotations; upstream terms in NOTICE
Checksums: per-config SHA-256 hashes in manifest.json
Configs: agent, chat, code, ifeval, math, serbian, uncertainty
No test split: create your own held-out split before evaluation

Citation

@misc{miniai2026quata2datav5,
  title        = {MiniAI Quata2 Data (v5 Clean)},
  author       = {{MiniAI}},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/M1n1A1/quata2-data-v5-clean}},
  note         = {Teacher-generated supervised fine-tuning data, cleaned release}
}

Acknowledgements

Thanks to the maintainers of WildChat-1M, OpenCodeInstruct, self-oss-instruct, code_contests, GSM8K, NuminaMath-1.5, CoCoNot, FineWeb-2, and Qwen3.8-27B.

Changelog

2026-10-29 — v5-clean: first cleaned release; 21,857 rows; seven configs; PII redaction; Wikipedia-grounded agent rows and credential-giveaway chat rows removed.

Maintainer

MiniAI (MiniAI)

Downloads last month: 25

Total file size:

45.2 MB