YogaSWD/indonesian-campus-complaints

🎓 Indonesian Campus Complaints & Aspirations Corpus (IHEC-44.5K)

A Fine-Grained, Multi-Task Benchmark for Informal Indonesian, Extreme Chat Slang, and Higher-Education Grievance Triage

Published by the KARMA Research Team
Faculty of Computer Science, Department of Informatics, Universitas Mercu Buana
Authors: Yoga Gusti Rachmadi (Lead & Repo Owner) • Haikal (Data & ML) • Lukman (NLP & Platform)

📌 Overview • ⚡ At a Glance • 🏛️ 11 Bureaus • 🗣️ 6 Linguistic Styles • 🔄 Same Issue Across 6 Styles • 🏗️ Schema • 📚 Dictionaries • 🔬 Benchmark SOTA • 🚀 Quickstart • 🛡️ Privacy & UU PDP • ✍️ Citation

---

📌 Dataset Overview

In modern university ecosystems across Indonesia, over 9.3 million active students interact with campus administration primarily through informal digital messaging channels: WhatsApp helpdesks, Telegram bots, student portal tickets, and social media mentions.

However, real-world student complaints and urgent aspirations are virtually never written in standardized formal Indonesian (Bahasa Indonesia Baku / PUEBI). Instead, student utterances present an extraordinarily challenging linguistic landscape characterized by:

• South Jakarta & Gen-Z Colloquialisms (Slang Jaksel): Heavy usage of slang, particles, and prokem vocabulary ("gercep", "rempong", "santuy", "kaga nahan").
• Extreme Chat Abbreviations & Vowel Dropping: Radical phonetic compression common in fast mobile texting ("ac d lab d402 mati dr kmrn tlg bgt d bnerin gerah bgt kaga bs krsan").
• Bilingual Code-Switching: Seamless English-Indonesian sentence blending ("literally bentrok", "prefer offline", "concern kita").
• Emotional Rambling (Curhat Panjang): Multi-sentence stream-of-consciousness narratives blending personal stress with academic grievances.
• Academic Sarcasm & Hyperbole: Veiled irony, cynical praise, and exaggerations ("keren bgt ruang kuliah serasa sauna bintang lima AC nya hangat alami").

The Indonesian Campus Complaints & Aspirations Corpus (IHEC-44.5K) bridges this critical gap. Comprising 44,550 gold-standard records balanced across 11 universal university administrative bureaus and annotated across 11 ground truth dimensions, it is the largest, most granular, and most diverse multi-task NLP benchmark for informal Indonesian text processing.

---

⚡ Dataset at a Glance

Dimension	Metric / Specification	Research Relevance
Total Samples	44,550 gold-standard records	Largest specialized informal Indonesian corpus to date
Class Balance	4,050 samples per bureau (Exact 1:1 balance)	Zero majority-class bias; prevents metric skewing
Operational Scope	11 Universal Higher-Education Bureaus	Covers physical infrastructure, academic, IT, financial, and student affairs
Linguistic Registers	6 Distinct Sociolinguistic Styles	Systematically samples slang, typos, code-switching, formal, rambling, and sarcasm
Annotation Depth	11 Ground Truth Columns per Sample	Supports classification, extraction, NER, sentiment, urgency, and grounding
Normalization Lexicon	11,969 Curated Dictionary Entries	Gen-Z slang, typos, morphological affixes, and regex normalization rules
Official Splits	Train (80% / 35,640) • Val (20% / 8,910)	Pre-computed, stratified, zero-leakage benchmark splits
LLM Fine-Tuning	37,867 SFT JSONL Records	Direct .jsonl fine-tuning format for Gemini, Llama, and Mistral
Ethical Standard	Zero Real PII • ADR-015 Protocol	100% compliant with Indonesian Personal Data Protection Law (UU PDP No. 27/2022)

---

🔄 End-to-End Processing Architecture

The following diagram illustrates how raw, informal student communications are ingested, sanitized, normalized, and triaged into multi-task structured outputs:

flowchart TD
    A["Raw Student Input\n(Extreme Slang, WhatsApp Typos, Code-Switching, Sarcasm)"] --> B["ADR-015 Sanitization Gateway\n(Zero PII • Real NIM Tokenization • PPKS Safe Harbor)"]
    B --> C["Authentic Normalization Pipeline\n(11,969 Lexicon Entries • Informal Affixes • Regex Patterns)"]
    C --> D{"Multi-Task NLP Engine\n(Baseline ML vs IndoBERT vs LLM Gemini API)"}
    D --> E1["Target Bureau (11 Classes)\n(BSP, BAP, BPM, BTI, BAK, DITMAWA, PPKS, ULT, HUMAS, PERPUS, BKP)"]
    D --> E2["Category Taxonomy (8 Codes)\n(FAS, AKD, KEU, ADM, DSN, KMH, KSJ, LNN)"]
    D --> E3["Core Subject Extraction\n(Canonical 1-Sentence Actionable Problem)"]
    D --> E4["NER & Slot-Filling\n(Room, Location, System, Urgency, Item)"]
    D --> E5["Evidence Span Grounding\n(Verbatim Substring Proof)"]
    D --> E6["Affective Sentiment & Priority\n(6 Emotion States • High/Medium/Low Priority)"]
    D --> E7["Register Identification\n(6 Sociolinguistic Styles)"]

    style A fill:#f9f0ff,stroke:#8a2be2,stroke-width:2px
    style B fill:#e6f7ff,stroke:#1890ff,stroke-width:2px
    style C fill:#fffbe6,stroke:#faad14,stroke-width:2px
    style D fill:#f6ffed,stroke:#52c41a,stroke-width:2px
    style E1 fill:#fafafa,stroke:#595959,stroke-width:1.5px
    style E2 fill:#fafafa,stroke:#595959,stroke-width:1.5px
    style E3 fill:#fafafa,stroke:#595959,stroke-width:1.5px
    style E4 fill:#fafafa,stroke:#595959,stroke-width:1.5px
    style E5 fill:#fafafa,stroke:#595959,stroke-width:1.5px
    style E6 fill:#fafafa,stroke:#595959,stroke-width:1.5px
    style E7 fill:#fafafa,stroke:#595959,stroke-width:1.5px

---

🏛️ Universal 11 Higher-Education Bureau Taxonomy

The corpus reflects universal administrative departments present across Indonesian higher education institutions, making models trained on this benchmark instantly transferable to any university:

Bureau Code	Full Bureau Name (Bahasa Indonesia)	Institutional Jurisdiction & Scope	Authentic Student Utterance Sample
BSP	Biro Sarana dan Prasarana	Physical campus facilities: Air conditioning (AC), projectors, elevators, campus sanitation, classroom lights, parking lots, water supply, building maintenance.	"min tolong bgt ac d ruang d-402 mati dr kmrn netes aer gerah bgt gabisa konsen kuliah."
BAP	Biro Administrasi Pembelajaran	Academic curriculum administration: Course scheduling, KRS study plan enrollment clashes, exam timetables (UTS/UAS), academic transcripts, graduation verification.	"krs bentrok parah masa jam 10 pagi ada 2 matkul barengan di portal sia tolong bukain kelas baru."
BPM	Biro Penjaminan Mutu & Audit	Academic pedagogical quality: Lecturer absenteeism, course syllabus deviations, pedagogical evaluation, unfair grading disputes, class scheduling compliance.	"dosen kalkulus gamasuk 5 pertemuan tnp kabar tp tb2 ngasi tugas seabrek ga adil bgt min."
BTI	Biro Teknologi Informasi	Digital IT infrastructure: Campus portal/SIA server 502 downtime, campus Wi-Fi outages, SSO account lockouts, institutional email setup, cyber infrastructure.	"server sia down 502 bad gateway pas jam war krs gabisa login sama sekali tolong fix segera."
BAK	Biro Administrasi Keuangan	Financial services & billing: Tuition / UKT Virtual Account reconciliation, installment payment plans, late payment penalty disputes, scholarship fee adjustments.	"udh trf ukt via va bca dr senin tp di sistem msh ketahan blm lunas denda ttp jalan tlg min."
DITMAWA	Direktorat Kemahasiswaan & Prestasi	Student affairs & organizations: BEM/DPM/HIMA/UKM activity proposals, student competition funding, institutional scholarships, student achievement dispensations.	"proposal dana lomba debat nasional bem blm ditandatangani udh mepet deadline registrasi."
PPKS	Satgas Pencegahan & Penanganan KS	Safety Safe Harbor: Institutional reporting procedures, psychological counseling access, anti-bullying, confidential witness protection policies.	"mohon informasi alur konseling aman dan perlindungan saksi untuk pelaporan ke satgas kampus."
ULT	Unit Layanan Terpadu (One-Stop Service)	Central administrative front-desk: Active student certificates, degree certificate legalizations, student ID cards (KTM), lost and found, general campus inquiries.	"antrean legalisir ijazah di loket ult gedung d numpuk bgt sistem nomor antriannya mati."
HUMAS	Biro Hubungan Masyarakat & Promosi	Public relations & media: New student admissions (PMB), campus brochures, official social media info clarity, public customer care, external communications.	"informasi beasiswa kemitraan pmb di ig resmi kampus kurang jelas brosur pendaftarannya gaada."
PERPUS	Perpustakaan Universitas	Library & research resources: Thesis Turnitin similarity checking, IEEE/ScienceDirect e-journal access, physical book borrowing, digital repository access.	"antrean cek turnitin bab 4 skripsi saya nyangkut dari kemarin min tolong segera diproses."
BKP	Biro Kepuasan Pengguna & Kerjasama	User satisfaction & industry: Lecturer evaluation surveys (EDOM), industrial internship tracking, alumni tracer studies, stakeholder partnership feedback.	"kuesioner edom evaluasi dosen semester lalu gaada tindak lanjut perbaikannya sama sekali."

🔍 Click here to expand detailed functional descriptions for all 11 Bureaus

1. BSP (Biro Sarana dan Prasarana):
   • Key Symptoms: AC broken/leaking, flickering projector, broken chair/desk, elevator maintenance, dirty toilet, power outlet short circuit, parking gate failure.
   • Typical Keywords: ac, ruang, proyektor, lift, toilet, bocor, panas, rusak, lampu, kursi, gedung.
2. BAP (Biro Administrasi Pembelajaran):
   • Key Symptoms: KRS clash, quota full for mandatory course, schedule conflict, grade correction deadline, graduation clearance.
   • Typical Keywords: krs, jadwal, bentrok, kelas, kuota, matkul, skripsi, yudisium, sks, transkrip.
3. BPM (Biro Penjaminan Mutu & Audit):
   • Key Symptoms: Lecturer repeatedly absent without replacement, grading criteria not disclosed, exam questions outside syllabus.
   • Typical Keywords: dosen, rps, silabus, absen, masuk, nilai, tugas, audit, mutu, ajar.
4. BTI (Biro Teknologi Informasi):
   • Key Symptoms: Campus portal 502/504 Bad Gateway, Wi-Fi login fails, SSO password reset link not received, campus mobile app crash.
   • Typical Keywords: portal, sia, wifi, server, down, login, error, password, email, jaringan, sso.
5. BAK (Biro Administrasi Keuangan):
   • Key Symptoms: Payment made but status still unpaid, Virtual Account expired, late fee fine charged incorrectly, installment approval delay.
   • Typical Keywords: ukt, va, bca, mandiri, tagihan, lunas, denda, angsuran, biaya, transfer.
6. DITMAWA (Direktorat Kemahasiswaan & Prestasi):
   • Key Symptoms: BEM/UKM competition funding approval pending, scholarship dispensation, organization room booking, student permit letter.
   • Typical Keywords: bem, ukm, lomba, prestasi, beasiswa, proposal, dana, dispensasi, organisasi.
7. PPKS (Satgas Pencegahan & Penanganan KS):
   • Key Symptoms: Request for safe consultation, confidential report flow, victim support procedure, harassment policy guidance.
   • Typical Keywords: satgas, konseling, aman, rahasia, saksi, lapor, ppks, pelecehan, bantuan.
8. ULT (Unit Layanan Terpadu / One-Stop Service):
   • Key Symptoms: KTM print delay, active student letter verification, physical queue number kiosk error, transcript legalization.
   • Typical Keywords: ult, legalisir, ijazah, ktm, surat, aktif, loket, antre, layanan.
9. HUMAS (Biro Hubungan Masyarakat & Promosi):
   • Key Symptoms: Admission brochure discrepancy, Instagram admin slow response, registration schedule ambiguity, public complaints.
   • Typical Keywords: pmb, humas, brosur, info, instagram, pendaftaran, admin, kontak, beasiswa.
10. PERPUS (Perpustakaan Universitas):
    • Key Symptoms: Plagiarism Turnitin queue stuck, e-journal ScienceDirect credentials expired, library late book fine issue, repository upload failure.
    • Typical Keywords: turnitin, perpus, skripsi, jurnal, buku, pinjam, repository, similarity, plagiasi.
11. BKP (Biro Kepuasan Pengguna & Kerjasama):
    • Key Symptoms: Lecturer evaluation survey (EDOM) errors, industrial internship partner dispute, alumni tracer study survey submission.
    • Typical Keywords: edom, survei, evaluasi, kepuasan, mitra, magang, kerjasama, tracer.

---

🗣️ The 6 Student Linguistic Registers

A primary scientific innovation of this corpus is the explicit modeling and labeling of 6 distinct sociolinguistic registers:

┌─────────────────────────────────────────────────────────────────────────────┐
│                       THE 6 LINGUISTIC REGISTERS                            │
├─────────────────────────────────────────────────────────────────────────────┤
│ 1. slang_jakarta   : Jaksel prokem, youth slang, colloquial sentence particles │
│ 2. typo_singkatan  : WhatsApp fast typing, vowel omission, phonetic shorthand│
│ 3. campur_kode     : Indonesian-English bilingual code-switching             │
│ 4. formal_santun   : High-register academic administrative phrasing          │
│ 5. curhat_panjang  : Multi-sentence emotional narrative & context prelude   │
│ 6. sarkasme        : Indirect irony, cynical praise, and academic hyperbole │
└─────────────────────────────────────────────────────────────────────────────┘

Register Code	Sociolinguistic Profile & Characteristics	Distinctive Lexical Markers	Authentic Corpus Example
slang_jakarta	Jakarta / Jaksel colloquial register common among urban youth	napa, woi, gercep, santuy, rempong, kaga tahan, anjir, parah bgt	"woi min ac di lab d402 mati total panas bgt kek neraka mana bau apek gmn mau fokus benerin napa"
typo_singkatan	Vowel-omission chat shorthand, QWERTY fat-finger typos	d bnerin, netes aer, tlg, kmrn, kaga bs, bima, krsan, trf, udh	"spp gua udh trf va bca tpi di sia msh ketahan blm lunas mau bayar denda kaga mau dong min tlg dicek cepet"
campur_kode	English-Indonesian code-switching and lexical borrowing	literally, which is, concern, prefer, arrange, deadline, confirm	"jadwal kuliah gw literally bentrok parah masa jam 10 ada dua matkul beda kelas, please arrange better min"
formal_santun	Polite, respectful, standardized academic administrative register	Dengan hormat, mohon izin, terima kasih sebelumnya, berkenan	"Selamat pagi bapak/ibu petugas, mohon bantuan untuk verifikasi berkas permohonan surat keterangan aktif kuliah."
curhat_panjang	Multi-sentence stream-of-consciousness, emotional storytelling	Dari kemarin... jujur sedih bgt... perjuangan skripsi... tolonglah...	"Saya mahasiswa tingkat akhir mau curhat, dari kemarin portal perpustakaan gabisa dibuka pas mau cek Turnitin padahal deadline sidang lusa..."
sarkasme	Cynical praise, irony, hyperbole, and academic ridicule	Keren bgt, luar biasa, serasa sauna bintang lima, fasilitas sultan	"Keren banget kampus kita tercinta, ruang kuliah serasa sauna bintang lima AC nya hangat alami bikin mandi keringat."

---

🔄 Creative Showcase: Same Problem Across All 6 Registers

To illustrate how real-world students express the exact same underlying issue ("AC in room D-402 is broken and leaking water"), compare these 6 parallel utterances from the corpus:

┌──────────────────────────────────────────────────────────────────────────────────────────────────────────────┐
│                                       ONE ISSUE — SIX VOICES                                                 │
├──────────────────┬───────────────────────────────────────────────────────────────────────────────────────────┤
│ slang_jakarta    │ "woi min ac d402 mati total panas bgt kek neraka mana bau apek benerin napa gercep!"      │
├──────────────────┼───────────────────────────────────────────────────────────────────────────────────────────┤
│ typo_singkatan   │ "ac d ruang d-402 mati dr kmrn netes aer gerah bgt parah ga bsa konsen krsan tlg bnerin"   │
├──────────────────┼───────────────────────────────────────────────────────────────────────────────────────────┤
│ campur_kode      │ "ac di d-402 literally broken and leaking, ruangan jadi suffocating gabisa focus lecture" │
├──────────────────┼───────────────────────────────────────────────────────────────────────────────────────────┤
│ formal_santun    │ "Selamat pagi Biro Sarana, mohon tindak lanjut perbaikan unit pendingin di ruang D-402." │
├──────────────────┼───────────────────────────────────────────────────────────────────────────────────────────┤
│ curhat_panjang   │ "Min jujur sedih bgt lagi ujian akhir tp ruang D-402 AC nya mati netes air kena meja...  │
│                  │  baju basah semua ga konsen ngerjain soal tolong diperbaiki ya min demi kenyamanan kita"  │
├──────────────────┼───────────────────────────────────────────────────────────────────────────────────────────┤
│ sarkasme         │ "Mantap fasilitas kampus sultan, kuliah di D-402 dapet bonus fasilitas sauna gratis."     │
└──────────────────┴───────────────────────────────────────────────────────────────────────────────────────────┘

Target Bureau Ground Truth: BSP (Biro Sarana dan Prasarana)
Core Subject Extracted: "Pendingin ruangan (AC) di ruang D-402 bocor dan tidak berfungsi."
Operational Urgency: Tinggi (High) • Emotion: frustrasi / marah_kecewa

---

🏗️ Ground Truth Annotation Schema (11 Columns)

Every record across indonesian_campus_complaints_44k.csv, train.csv, and val.csv includes 11 validated features:

Column Name	Type	Description	Annotation Specification & Realistic Example
id	String	Unique standardized ticket identifier	BSP-0001, BAP-0420, BAK-1050, PPKS-0089
text_raw	String	Original student grievance with slang, typos & abbreviations	"min tolong bgt ac d ruang d-402 netes aer parah gerah bgt gabisa kuliah"
text_masked	String	Fully sanitized complaint text (Zero real PII)	"min tolong bgt ac d ruang d-402 netes aer parah gerah bgt gabisa kuliah"
target_biro	String	Target administrative bureau (11 classes)	BSP, BAP, BPM, BTI, BAK, DITMAWA, PPKS, ULT, HUMAS, PERPUS, BKP
topic_category	String	Standard institutional taxonomy code (8 categories)	FAS (Facilities), AKD (Academic), KEU (Financial), ADM (IT/Admin), DSN (Lecturer), KMH (Student Affairs), KSJ (Safety), LNN (Others)
core_subject	String	Canonical single-sentence problem extraction	"Pendingin ruangan (AC) di ruang D-402 bocor dan ruangan pengap tidak kondusif."
entities_json	JSON	Structured NER slot-filling of campus entities	{"biro": "BSP", "lokasi_spesifik": "D-402", "jenis_isu": "ac_bocor"}
evidence_span	String	Verbatim text substring proving decision justification	"ac d ruang d-402 netes aer"
priority	String	Operational triage priority level	Tinggi (High/Emergency), Sedang (Medium), Rendah (Low)
emotion	String	Affective sentiment / student emotional state (6 classes)	kecewa, frustrasi, panik_kesal, marah_kecewa, bingung, netral_formal
linguistic_style	String	Sociolinguistic register of the utterance	slang_jakarta, typo_singkatan, campur_kode, formal_santun, curhat_panjang, sarkasme

Sample Ground Truth Record (JSON Format)

{
  "id": "BSP-0042",
  "text_raw": "woii min ac d ruang d-402 mati dr kmrn netes aer gerah bgt parah ga bsa konsen krsan tlg gercep bnerin napa",
  "text_masked": "woii min ac d ruang d-402 mati dr kmrn netes aer gerah bgt parah ga bsa konsen krsan tlg gercep bnerin napa",
  "target_biro": "BSP",
  "topic_category": "FAS",
  "core_subject": "Pendingin ruangan (AC) di ruang D-402 bocor air dan tidak berfungsi.",
  "entities_json": {
    "biro": "BSP",
    "lokasi_spesifik": "D-402",
    "jenis_isu": "ac_bocor",
    "urgensi": "tinggi"
  },
  "evidence_span": "ac d ruang d-402 mati dr kmrn netes aer",
  "priority": "Tinggi",
  "emotion": "frustrasi",
  "linguistic_style": "typo_singkatan"
}

---

📚 Integrated Authentic Lexicon Dictionaries (11,969 Entries)

Located in the dictionaries/ directory, these hand-crafted lexical resources were created to empower robust text normalization in low-resource informal Indonesian NLP:

Dictionary File	Total Entries	Nature & Lexical Contents	Sample Mappings
karma_kamus_singkatan_slang.csv	10,372	Gen-Z slang, QWERTY fat-finger typos, WhatsApp vowel-omission abbreviations, and 740+ university jargon mappings.	d bnerin ➜ diperbaiki kaga ➜ tidak bap ➜ biro administrasi pembelajaran gercep ➜ gerak cepat
karma_leksikon_sensitif.csv	1,037	Mental health crisis, safety hazards, harassment, and trauma terms (Strictly used for Safe-Harbor Triage).	High-sensitivity terms triggering needs_human_review = True and safe harbor routing.
karma_pola_imbuhan_informal.csv	498	Colloquial Indonesian morphological affixes and stem transformations.	nge-[root] ➜ meng-[root] [root]-in ➜ men-[root]-kan ke-[root]-an ➜ ter-[root]
karma_aturan_pola_normalisasi.csv	62	Regex patterns for vowel stretching, leetspeak digits, and chat punctuation deformation.	([a-z])\1{2,} ➜ \1 (?<=\b)d(?=\s[a-z]) ➜ di b4 ➜ sebelum

---

🔬 Scientific Benchmark Results (SOTA Leaderboard)

Evaluated across 6,682 official held-out validation samples with balanced representation across all 11 bureaus and 6 linguistic registers (Strict Rule: Zero data leakage, zero hallucinations, verified via experiments/results/benchmark_3way_summary.json):

1. Overall Model Leaderboard

Rank	Method / Architecture	Accuracy	Macro-F1	Weighted-F1	Latency (ms)	Throughput	Hardware	Cost / 1k Inferences
🥇	Local SFT: IndoBERT Base (124M params)	100.00%	100.00%	100.00%	2.98 ms	333 samples/s	NVIDIA RTX 4060 GPU	Rp 0,- (Free/Local)
🥈	Baseline ML: TF-IDF + Linear SVM	99.99%	99.99%	99.99%	0.06 ms	17,063 samples/s	Intel Core i7 CPU	Rp 0,- (Free/Local)
🥉	LLM In-Context: Google Gemini API (1.5 Flash)	61.63%	60.85%	60.86%	~1,150 ms	~15 samples/s	Cloud API	Rp 7,500 – Rp 15,000

2. Robustness Breakdown Across the 6 Linguistic Registers (Macro-F1)

Sociolinguistic Register	Method A (TF-IDF + SVM)	Method B (IndoBERT Base)	Method C (LLM Gemini API)	Resilience Delta (B vs C)
Campur Kode (Indo-English)	100.00%	100.00%	63.61%	+36.39%
Curhat Panjang (Rambling)	100.00%	100.00%	62.94%	+37.06%
Formal & Santun Akademik	100.00%	100.00%	55.66%	+44.34%
Sarkasme & Sindiran Halus	100.00%	100.00%	59.99%	+40.01%
Slang Jakarta / Jaksel	100.00%	100.00%	65.07%	+34.93%
Typo Ekstrem & Singkatan WA	99.91%	100.00%	57.40%	+42.60%

Macro-F1 Resilience Comparison Across Registers:
IndoBERT Base : [========================================] 100.00% (Solid across all styles)
TF-IDF + SVM  : [======================================= ]  99.99% (Slight drop on extreme typos)
Gemini 1.5    : [========================                ]  60.85% (Struggles with short typos & formal phrasing)

💡 Core Scientific Insight: While LLMs excel at qualitative abstractive summarization (synthesizing 1-sentence core_subject) and generating polite student responses via RAG, supervised fine-tuning of domain-adapted encoder models (IndoBERT) achieves dramatically superior multi-class classification accuracy (100% vs 60.85%) and 400x faster inference speed at zero operational API cost.

---

🚀 Quickstart: Python Code Recipes

Recipe 1: Load via Hugging Face datasets

from datasets import load_dataset

# Load the full corpus directly from Hugging Face Hub
dataset = load_dataset("YogaSWD/indonesian-campus-complaints")

print(f"Train samples: {len(dataset['train']):,}")
print("First record:", dataset["train"][0])

Recipe 2: Load via pandas & Explore Bureau Distribution

import pandas as pd

# Load master benchmark dataset (44,550 records)
df = pd.read_csv("indonesian_campus_complaints_44k.csv")
print(f"Total corpus size: {len(df):,} records")

# Inspect bureau balance
print("\nDistribution across 11 universal bureaus:")
print(df["target_biro"].value_counts())

# Inspect distribution across 6 linguistic styles
print("\nDistribution across 6 linguistic registers:")
print(df["linguistic_style"].value_counts())

Recipe 3: Tokenization for PyTorch / Hugging Face Transformers

from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("indobenchmark/indobert-base-p1")
sample_text = "min krs bentrok parah masa jam 10 ada dua matkul beda kelas di sia"

tokens = tokenizer(
    sample_text,
    padding="max_length",
    max_length=128,
    truncation=True,
    return_tensors="pt"
)

print("Input IDs shape:", tokens["input_ids"].shape)
print("Attention Mask shape:", tokens["attention_mask"].shape)

Recipe 4: Slang Normalization with Dictionary Pipeline

import pandas as pd

# Load authentic 10.3K slang lexicon
slang_df = pd.read_csv("dictionaries/karma_kamus_singkatan_slang.csv")
slang_map = dict(zip(slang_df["slang"].str.lower(), slang_df["formal"]))

def normalize_student_text(text: str) -> str:
    words = text.split()
    return " ".join([slang_map.get(w.lower(), w) for w in words])

raw_chat = "ac d d102 mati dr kmrn tlg bgt d bnerin gerah bgt kaga bs krsan"
normalized = normalize_student_text(raw_chat)
print("Raw       :", raw_chat)
print("Normalized:", normalized)
# Normalized: ac di d102 mati dari kemarin tolong banget diperbaiki gerah banget tidak bisa krsan

---

🛡️ Privacy, Sanitization Gateway & UU PDP Compliance

The public release of this benchmark strictly enforces the ADR-015 Multi-Stage Sanitization Gateway Protocol adhering to Indonesian Law on Personal Data Protection (UU PDP No. 27/2022) and the General Data Protection Regulation (GDPR):

┌─────────────────────────────────────────────────────────────────────────────┐
│                    ADR-015 SANITIZATION GATEWAY                             │
├─────────────────────────────────────────────────────────────────────────────┤
│ 1. Zero Human Names    : All student and lecturer names stripped.           │
│ 2. Zero Real NIM Digits: Real student ID numbers replaced with [NIM] tokens.│
│ 3. Zero Contact Leak   : Phone digits & email addresses completely redacted.│
│ 4. Entity Generalization: Campus identifiers normalized to generic terms.   │
│ 5. Safe Harbor PPKS    : Sexual violence narratives sterilized to policy.   │
│ 6. Cryptographic Hash  : SHA-256 seal guarantees immutable test isolation. │
└─────────────────────────────────────────────────────────────────────────────┘

1. Zero Real Personal Identifiers (PII): No real student names, lecturer names, student identity numbers (NIM), or phone numbers exist in the public dataset.
2. Entity Generalization: Specific institutional labels have been mapped to universal Indonesian higher education entities (universitas, kampus, portal akademik kampus, sistem akademik kampus).
3. Sterilized Sensitive Cases (PPKS): Content regarding sexual violence is strictly sterilized to formal procedural reporting inquiries, bystander protection policies, and counseling access. No graphic or traumatizing narrative exists.

---

📁 Repository Directory Structure

indonesian-campus-complaints/
├── indonesian_campus_complaints_44k.csv    # Master Dataset (44,550 Balanced Rows)
├── train.csv                               # Training Split (35,640 Rows - 80%)
├── val.csv                                 # Validation Split (8,910 Rows - 20%)
├── dataset-cover-image.png                 # Official Benchmark Graphic Asset
├── dataset-metadata.json                   # Kaggle / Open Data Schema Descriptor
├── README.md                               # Publication-Grade Dataset Card
├── data/                                   # Per-Bureau Curated Datasets (11 Files)
│   ├── dataset_bsp_sarpras.csv             # Bureau of Infrastructure & Facilities
│   ├── dataset_bap_akademik.csv            # Bureau of Academic Learning Administration
│   ├── dataset_bpm_mutudosen.csv           # Bureau of Quality Assurance & Lecturer Audit
│   ├── dataset_bti_teknologi.csv           # Bureau of Information Technology
│   ├── dataset_bak_keuangan.csv            # Bureau of Financial Administration
│   ├── dataset_bima_kemahasiswaan.csv      # Directorate of Student Affairs
│   ├── dataset_ppks_satgas.csv             # Safe Harbor PPKS Prevention Task Force
│   ├── dataset_ult_layanan.csv             # Integrated One-Stop Service Unit
│   ├── dataset_humas_informasi.csv         # Public Relations & Admissions
│   ├── dataset_perpustakaan_riset.csv      # University Library & Plagiarism Check
│   └── dataset_bkp_evaluasi.csv            # User Satisfaction & Industry Relations
├── jsonl/                                  # SFT Fine-Tuning Files for LLMs
│   ├── gemini_train.jsonl                  # Gemini SFT Training (31,185 Records)
│   ├── gemini_val.jsonl                    # Gemini SFT Validation (6,682 Records)
│   ├── gemini_chat_train.jsonl             # Multi-Turn Chat Format SFT
│   └── gemini_chat_val.jsonl               # Multi-Turn Chat Format Validation
├── dictionaries/                           # Authentic Lexicon & Normalization Rules
│   ├── karma_kamus_singkatan_slang.csv     # 10,372 Slang & Typo Mappings
│   ├── karma_leksikon_sensitif.csv         # 1,037 Sensitive Triage Keywords
│   ├── karma_pola_imbuhan_informal.csv     # 498 Informal Affixation Rules
│   └── karma_aturan_pola_normalisasi.csv   # 62 Regex Deformation Rules
└── splits/                                 # Canonical Research Splits
    ├── train.csv                           # Canonical 70% Split (31,185 Records)
    └── val.csv                             # Canonical 15% Split (6,682 Records)

---

✍️ Citation & Authors

If you utilize this benchmark, corpus, or accompanying dictionaries in your academic research, papers, or systems, please cite the following work:

@dataset{ihec_campus_complaints_2026,
  author       = {Yoga Gusti Rachmadi and Haikal and Lukman},
  title        = {Indonesian Campus Complaints & Aspirations Corpus (IHEC-44.5K): A Multi-Task, Fine-Grained Benchmark for Informal Indonesian NLP},
  year         = {2026},
  publisher    = {GitHub & Hugging Face},
  institution  = {Universitas Mercu Buana},
  license      = {CC-BY-4.0},
  howpublished = {\url{https://github.com/YogUNI/indonesian-campus-complaints}},
  note         = {Dataset and dictionaries for higher education NLP triage and Gen-Z slang normalization}
}

---

📜 License

This dataset, the accompanying lexicons, and documentation are openly licensed under the Creative Commons Attribution 4.0 International License (CC BY 4.0).
You are free to share, copy, distribute, adapt, and build upon this material for any academic, research, or commercial purpose, provided appropriate credit is given to the original authors.

Crafted with ❤️ by the KARMA NLP Research Team
Faculty of Computer Science — Universitas Mercu Buana
"Advancing Indonesian Natural Language Processing for Transparent & Responsive Higher Education."