Daffaadityp/smkn26-chat-sft-id-en

SMKN 26 Jakarta — PEFT Chat Dataset

Dataset percakapan bilingual (Indonesia + English) untuk fine-tuning asisten resmi SMK Negeri 26 Jakarta memakai PEFT (LoRA / DoRA / QLoRA).

Dirancang agar model:

• menjawab fakta sekolah dengan akurat
• menangani pertanyaan orang tua / calon siswa secara natural
• menolak mengarang saat data tidak tersedia (anti-hallucination)

Source of truth: Dataset SMKN 26 Jakarta.pdf → curated KB kb/smkn26_facts.json
Rebuild: python scripts/build_dataset.py

---

Snapshot

Item	Value
Total conversations	2,010
Parallel pairs	1,005 (id+en)
train / validation / test	1,710 / 200 / 100
Format	Chat messages (system → user → assistant)
Languages	Parallel id / en (twins share split)
Persona	Official SMKN 26 Jakarta assistant
Avg user turns	~1.12 (includes multi-turn dialogues)
Leakage control	Bilingual twins always share the same split

Category mix (latest build)

Category	Count	Share
jurusan	338	16.8%
identitas	220	11.0%
kurikulum_pkl	218	10.9%
sejarah	200	10.0%
ppdb_alumni	188	9.4%
lsp_blud	182	9.1%
prestasi	176	8.8%
fasilitas	164	8.2%
multi_turn	120	6.0%
visi_misi	120	6.0%
anti_hallucination	84	4.2%

Exact machine-readable stats: data/build_report.json.

Why this dataset is PEFT-friendly

1. Grounded answers — every non-refusal sample traces to fact_ids in the KB
2. Natural phrasing — multiple answer styles (concise / explanatory) without inventing facts
3. Scenario & multi-turn — realistic parent/student flows, not only single-fact quizzes
4. Anti-hallucination — explicit refusal training for unknown/out-of-domain questions
5. Balanced categories — category-aware sampling toward the design mix

---

Categories

Category	Focus
identitas	NPSN, address, contacts, accreditation, ISO
sejarah	1971 founding → STM Pembangunan → SMKN 26 (1997)
visi_misi	Vision, mission, motto, slogan
jurusan	KGSP, TEDK, TTL, TFLM, TMPO, SIJA
kurikulum_pkl	4-year system, Kurikulum Merdeka, internship rules
lsp_blud	LSP-P1 / BNSP schemes, BLUD & Teaching Factory units
fasilitas	Labs, auditorium, transport access, Adiwiyata
ppdb_alumni	Admission quotas, requirements, IASPEM26
prestasi	LKS, KIR, Bionic26 / ASCC, alumni abroad
multi_turn	2–4 turn dialogues
anti_hallucination	Missing/unknown/out-of-scope → refuse

---

Data fields

Field	Type	Description
id	string	Unique sample id ({pair_id}_{lang})
pair_id	string	Shared id for ID↔EN twins
lang	id | en	Language of the conversation
category	string	Content bucket
fact_ids	string[]	KB fact references (empty/unknown-linked for refusals)
split	train/validation/test	Dataset split
difficulty	easy/medium/hard	Rough difficulty label
turns	int	Number of user turns
messages	list	Chat messages for SFT trainers

Example

{
  "id": "smkn26_fact_0001_id",
  "pair_id": "smkn26_fact_0001",
  "lang": "id",
  "category": "identitas",
  "fact_ids": ["fact.identitas.npsn"],
  "split": "train",
  "difficulty": "easy",
  "turns": 1,
  "messages": [
    {
      "role": "system",
      "content": "Anda adalah asisten resmi SMK Negeri 26 Jakarta. Jawab akurat berdasarkan pengetahuan sekolah. Jika data tidak tersedia, katakan tidak tahu — jangan mengarang."
    },
    {
      "role": "user",
      "content": "Apa NPSN SMK Negeri 26 Jakarta?"
    },
    {
      "role": "assistant",
      "content": "Untuk pertanyaan tentang NPSN: 20103787"
    }
  ]
}

---

Quickstart

Load with 🤗 Datasets

from datasets import load_dataset

ds = load_dataset(
    "json",
    data_files={
        "train": "data/train.jsonl",
        "validation": "data/validation.jsonl",
        "test": "data/test.jsonl",
    },
)
print(ds)
print(ds["train"][0]["messages"])

Use with TRL SFT (chat messages)

from datasets import load_dataset
from trl import SFTTrainer  # example; configure model/tokenizer yourself

ds = load_dataset("json", data_files={"train": "data/train.jsonl", "validation": "data/validation.jsonl"})
# Pass formatting_func / chat template that reads sample["messages"]

Rebuild locally

python scripts/build_dataset.py
python -m pytest tests/ -v

---

Quality controls

• Required schema fields validated in tests/test_dataset_quality.py
• Non-refusal answers must reference KB facts with confidence != unknown
• Refusal samples must not invent school facts
• Unique ids; bilingual pair co-split enforced
• Split ratios kept near 85 / 10 / 5

---

Intended use

• SFT / PEFT for a school information assistant about SMKN 26 Jakarta
• Offline eval of factual accuracy + refusal behavior on test

Out of scope

• Live tuition fees, daily schedules, personal student/staff data
• General knowledge unrelated to SMKN 26
• Replacing official school announcements

Limitations

• Grounded only on the curated PDF-derived KB
• Some PDF fields are incomplete (e.g. exact GPS) → marked unknown and used for refusal training
• Principal naming may be partial when sources disagree
• Synthetic paraphrases increase coverage but are not human-annotated dialogues

---

License

Creative Commons Attribution 4.0 International (CC BY 4.0)

You may copy, redistribute, remix, transform, and build upon this dataset for any purpose (including commercial fine-tuning), as long as you give appropriate credit.

This dataset was curated from publicly available open-web information about SMK Negeri 26 Jakarta (aggregated/researched from open sources, then structured into bilingual chat samples). The license covers this curated dataset release; it does not change the terms of any original third-party web pages that were consulted.

Suggested attribution:

SMKN 26 Jakarta PEFT Chat Dataset (CC BY 4.0). Curated from publicly available information about SMK Negeri 26 Jakarta.

Full legal text: https://creativecommons.org/licenses/by/4.0/

Citation

@misc{smkn26_peft_chat_2026,
  title        = {SMKN 26 Jakarta PEFT Chat Dataset},
  author       = {Daffa Aditya Pratama},
  year         = {2026},
  howpublished = {Hugging Face Datasets},
  note         = {Bilingual chat SFT dataset curated from open-web public information about SMK Negeri 26 Jakarta},
  license      = {CC-BY-4.0}
}

Please credit SMK Negeri 26 Jakarta as the institutional subject of the knowledge.