kodetr/sasak-benchmark-100k

🌴 SasakNLP: Morphological Benchmark & Authentic Sasak Corpus

Official research dataset accompanying the SasakNLP toolkit, developed by kodetr (kodetr.com). Bahasa Sasak is an Austronesian low-resource regional language spoken by approximately 3 million people across Lombok, West Nusa Tenggara (NTB), Indonesia.

• Website Resmi: https://kodetr.com
• Repositori GitHub: https://github.com/kodetr/sasaknlp
• Web Demo Space: https://huggingface.co/spaces/kodetr/sasaknlp-demo

📂 Dataset Subsets (Configs)

1. benchmark (benchmark_100k.parquet & .csv): 100,000 verified morphological test pairs (surface, lemma, prefix, infix, suffix, dialect) curated following Sastrawi-grade purity standards.
2. corpus (sasak_sentences_large.parquet & .csv): 12,591 authentic sentences collected from folklore (Putri Mandalika, Dewi Anjani, Datu Doyan Nada), Balai Bahasa NTB examples, and verified regional publications.
3. lexicon (kamus_balai_bahasa_ntb.parquet & .csv): 2,761 official lexicon entries harvested from Balai Bahasa Provinsi NTB (Kamus Terpadu Sasambo).

🚀 Usage with Python datasets

from datasets import load_dataset

# 1. Load the benchmark dataset (100,000 pairs)
benchmark = load_dataset("kodetr/sasak-benchmark-100k", "benchmark")
print(benchmark['train'][0])

# 2. Load authentic sentences corpus (12,591 sentences)
corpus = load_dataset("kodetr/sasak-benchmark-100k", "corpus")
print(corpus['train'][0])

# 3. Load Balai Bahasa NTB dictionary
lexicon = load_dataset("kodetr/sasak-benchmark-100k", "lexicon")
print(lexicon['train'][0])

🛠 Accompanying Toolkit

Install the official Python library:

pip install sasaknlp

Quick example:

from sasaknlp import SasakNLP

nlp = SasakNLP()
res = nlp.process("Tiyang mangan nasiq bareng-bareng kance baturne.")
print(res.lemmas)

📊 Scientific Benchmark Figures & Dataset Architecture

The benchmark datasets were curated and evaluated through a rigorous empirical pipeline:

Figure 5: 3-Phase Dataset Acquisition & Curation Protocol

Figure 5: 3-phase scientific protocol: multi-source authentic ingestion (NTB folklore, linguistic corpora, Balai Bahasa dictionary), 5-stage cleaning & Sastrawi-grade curation, and generation of the 100k gold-standard benchmark datasets.

Figure 1: Morphological Disambiguation Accuracy by Affix Class

Figure 1: Morphological accuracy breakdown across prefixes, suffixes, infixes, circumfixes, and reduplication evaluated on the 100,000-pair benchmark.

Figure 2: Empirical Performance Across 5 Sasak Dialect Clusters

Figure 2: Cross-dialect morphological disambiguation accuracy and F1-score across the five major Sasak dialect clusters (Selaparang, Ngeno-Ngene, Mriak-Mriku, Ngeto-Ngete, and Kuto-Kute).

Figure 3: Error Taxonomy Distribution & Failure-Mode Diagnosis

Figure 3: Error taxonomy distribution (donut chart) and failure-mode diagnostic matrix demonstrating 87.98% exact lemma extraction, 11.92% Out-of-Vocabulary fallback, and near-zero overstemming (0.00%).

Figure 4: Computational Throughput & Benchmark Latency

Figure 4: High-throughput processing benchmark (22,500 words/second) and sub-millisecond average per-token latency (0.044 ms) with flat scaling curves across varying sequence lengths.

---

📖 Citation

@software{sasaknlp2026,
  author    = {kodetr},
  title     = {SasakNLP: Research-Grade Natural Language Processing and Morphological Toolkit for Bahasa Sasak},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://github.com/kodetr/sasaknlp}
}