themes12/lanna-global-mmlu-lite

If you use this dataset, please feel free to contact me by opening an issue or sending me a direct message. I would appreciate any feedback and would be happy to connect with others working on related topics. Thank you!

Lanna Global MMLU Lite 🏔️

A small, hand-validated multiple-choice benchmark dataset in Lanna script (ᨲᩫ᩠ᩅᨾᩮᩬᩥᨦ / Tua Mueang / Tai Tham) — the traditional writing system of the ancient Lanna Kingdom of northern Thailand.

This dataset is a Lanna-script translation of a subset of Global MMLU Lite.

⚠️ Use at your own risk. This is a personal research project. Translations were done by hand to the best of my ability and cross-checked against dictionaries and academic resources, but mistakes are expected.

---

What Is Lanna Script?

Lanna (also called Tua Mueang, Tai Tham, or Northern Thai script) is an abugida script historically used throughout the ancient Lanna Kingdom (present-day northern Thailand, Shan State of Myanmar, Laos, and Yunnan, China). It is still used in Buddhist religious contexts and by some cultural communities. It is a genuinely low-resource script with virtually no LLM training data.

---

Dataset Summary

Property	Value
Questions	31 (hand-validated)
Answer choices	4 (A / B / C / D)
Script	Lanna / Tai Tham (ᨲᩫ᩠ᩅᨾᩮᩬᩥᨦ)
Source	Global MMLU Lite
Translation	English → Thai (ChatGPT) → Lanna (Gemini) → human-validated
License	CC BY 4.0

---

Dataset Structure

Splits

Split	# Examples
test	31

Fields

Field	Type	Description
sample_id	string	Original Global MMLU Lite ID, e.g. high_school_geography/test/27
subject	string	Subject name, e.g. high_school_geography
subject_category	string	Broad category: STEM, Humanities, Social Sciences, Business, Medical, Other
split	string	Dataset split from source (test)
question_index	int	Question index within source subject
question	string	Question text in Lanna script
option_a	string	Option A in Lanna script
option_b	string	Option B in Lanna script
option_c	string	Option C in Lanna script
option_d	string	Option D in Lanna script
answer	string	Correct answer letter: A, B, C, or D
original_question	string	Original English question text
original_option_a	string	Original English option A
original_option_b	string	Original English option B
original_option_c	string	Original English option C
original_option_d	string	Original English option D
cultural_sensitivity_label	string	CA (Culturally Agnostic) or CS (Culturally Sensitive)
culture	string	Serialized list of cultures referenced in the question
region	string	Serialized list of geographic regions referenced
country	string	Serialized list of countries referenced
required_knowledge	string	Per-option knowledge type required (none, regional, cultural, dialect)
time_sensitive	string	Whether each option contains time-sensitive information
translator	string	Translation provenance (llm-human = LLM + human correction)
translation_status	string	human = hand-validated

---

Subject Coverage

Questions span 6 subject categories across 23 topics from the MMLU taxonomy:

Category	Subjects
Business	Business Ethics, Management, Professional Accounting
Humanities	High School European History, Prehistory, Logical Fallacies, Moral Disputes, Jurisprudence
Medical	College Medicine, Human Aging, Nutrition
Other	Global Facts, Miscellaneous
STEM	Astronomy, High School Statistics, High School Physics
Social Sciences	High School Geography, Sociology, Professional Psychology, Public Relations

---

Mapping back to Global MMLU Lite

Each sample_id maps directly to the source dataset:

---

LLM Evaluation Results

Evaluated zero-shot (no system prompt, no Lanna-specific context) on 23 questions:

Model	English Accuracy	Lanna Accuracy	Δ Drop
google/gemini-3.1-pro-preview	91.3%	69.6%	−21.7%
openai/gpt-5.4	87.0%	60.9%	−26.1%
anthropic/claude-opus-4.6	87.0%	61.9%	−25.1%
deepseek/deepseek-v3.2	91.3%	36.4%	−54.9%
google/gemini-3.1-flash-lite-preview	82.6%	47.8%	−34.8%
qwen/qwen3.5-9b	81.0%	23.8%	−57.1%

Key observations:

• All models perform significantly worse on Lanna than on English — as expected for a truly low-resource script.
• Gemini 3.1 Pro performs best on Lanna (69.6%), suggesting broader multilingual training.
• DeepSeek and Qwen show dramatic drops, suggesting strong English reasoning that doesn't generalize to unseen scripts.
• Models take 3–10× longer to process Lanna questions, suggesting real character-level decoding rather than pattern-matching.

⚠️ Sample size is only 23 questions — interpret differences cautiously.

---

Translation Methodology

1. English → Thai — Translated using ChatGPT
2. Thai → Lanna — Translated using Gemini (3.1 Pro Preview) with chain-of-thought prompting that provided grammar conventions and Lanna script rules
3. Human validation — All questions were corrected and validated by hand against Lanna dictionaries and academic guides

The translator is a native Northern Thai speaker. Therefore, I am confident that the tones, meanings, and cultural authenticity have been represented to the best of my ability. However, the written form may contain errors, as I am not fully proficient in Lanna script. This is partly because Lanna script is rarely taught within the compulsory education system, and only a very small number of compulsory education schools currently teach it.

---

Limitations

• Small sample size — 31 questions evaluated; don't over-interpret small differences.
• Translation quality — Hand-validated by a non-expert but native speaker. Lanna orthography conventions vary by region and tradition as there are still little standard writing conventions.
• No test-train leakage check — Original Global MMLU Lite questions may appear in model training data.
• No special prompting — Results reflect zero-shot performance with no Lanna-specific context given to the model.

---

Citation

If you use this dataset, please also cite the original Global MMLU Lite:

@dataset{GlobalMMLULite,
  title  = {Global MMLU Lite},
  author = {CohereForAI},
  url    = {https://huggingface.co/datasets/CohereForAI/Global-MMLU-Lite}
}

---

License

Translations in this dataset are released under CC BY 4.0.
The underlying questions are from Global MMLU Lite — please check their license for usage terms.

---

See Also

• 📝 Lanna and LLM Investigation — t3m.dev — Blog post with more context on this investigation
• 🤗 Global MMLU Lite (CohereForAI) — The original benchmark
• 💻 GitHub Repository — No source code and evaluation script provided as the research is in very beginning stage.