
Seni
Benchmark profile
Each underlying Malaysian cultural knowledge item is instantiated in parallel MCQ, CSRQ, and CRQ formats. This holds the knowledge target constant while changing answer exposure and scoring.
Coverage
MyCulture is a controlled testbed, not a claim to cover every Malaysian identity, community, or tradition.

Seni

Pakaian

Adat

Hiburan

Makanan

Agama
Parallel format example
The example below assesses knowledge about the traditional Lansaran game of the Murut community.
Which statements about the Lansaran game and the culture of the Murut community are correct?
Which statements about the Lansaran game and the culture of the Murut community are correct?
Describe the physical features, construction method, and community role in the traditional Lansaran game of the Murut community, based on the traditional practices described.
Full model performance
The complete MyCulture result table becomes an analytical surface: filter model families, rank any metric, and draw direct score profiles across MCQ, CSRQ, and CRQ.
Instrument score profile
Full published results
| Plot | Family | Model | Reasoning | |||||
|---|---|---|---|---|---|---|---|---|
| OpenAI | GPT-5 | Yes | 88.14 | 78.10 | 56.78 | 68.41 | 22.90 | |
| OpenAI | GPT-4o | — | 90.28 | 78.00 | 56.01 | 68.61 | 21.14 | |
| Gemini-2.5-Pro | Yes | 85.56 | 77.78 | 55.68 | 66.87 | 20.26 | ||
| OpenAI | GPT-5-mini | Yes | 84.73 | 77.50 | 55.46 | 67.25 | 25.65 | |
| Z.ai | GLM-4.6 | — | 87.37 | 77.45 | 54.91 | 56.72 | 17.68 | |
| AI SG | Qwen-SEA-LION-v4-32B-IT | — | 83.25 | 76.16 | 53.54 | 62.55 | 18.23 | |
| Z.ai | GLM-4.6 | Yes | 84.46 | 76.37 | 53.38 | 63.20 | 17.74 | |
| Alibaba | Qwen3-30B-Instruct | — | 82.48 | 75.21 | 52.50 | 59.03 | 17.13 | |
| Alibaba | Qwen3-235B-Instruct | — | 80.12 | 73.55 | 51.29 | 62.93 | 18.34 | |
| OpenAI | GPT-4o-mini | — | 78.36 | 74.30 | 49.20 | 65.78 | 20.37 | |
| OpenAI | GPT-5-nano | Yes | 80.12 | 73.61 | 48.38 | 63.10 | 20.76 | |
| Gemini-2.5-Flash-Lite | Yes | 79.57 | 72.90 | 48.00 | 62.56 | 18.95 | ||
| Z.ai | GLM-4-32B-128K | — | 75.18 | 71.88 | 47.83 | 59.26 | 16.09 | |
| Gemini-2.5-Flash | Yes | 82.87 | 71.05 | 47.39 | 65.23 | 21.47 | ||
| Alibaba | Qwen3-30B-Thinking | Yes | 79.57 | 70.49 | 43.82 | 62.48 | 17.35 | |
| Alibaba | Qwen3-235B-Thinking | Yes | 61.01 | 62.62 | 42.12 | 50.13 | 18.51 | |
| Alibaba | Qwen3-4B-Thinking | Yes | 71.61 | 66.42 | 37.67 | 48.47 | 10.98 | |
| Alibaba | Qwen3-4B-Instruct | — | 76.06 | 65.64 | 35.48 | 55.97 | 15.27 | |
| SeaLLM | SeaLLMs-v3-7B | — | 29.98 | 52.96 | 26.30 | 47.62 | 7.86 | |
| Baseline | Random baseline | — | 25.53 | 40.33 | 4.28 | — | — | |
| SeaLLM | SeaLLMs-v3-1.5B* | — | 0.00 | 0.16 | 0.00 | 39.54 | 6.45 | |
| Mesolitica | MaLLaM-1.1B* | — | 0.00 | 0.00 | 0.00 | 39.56 | 5.44 | |
| Mesolitica | MaLLaM-5B* | — | 0.00 | 0.00 | 0.00 | 46.10 | 6.64 |
Construction and validation
Reliable Malaysian cultural materials are transformed into matched evaluation items through automated checks and a qualified-resident audit loop.
JKKN materials and relevant academic publications.
Draft MCQ, CSRQ, and CRQ versions from shared evidence.
Check format compliance, answerability, and evidence consistency.
Strong models flag potentially ill-posed items for review.
Five qualified Malaysian residents audit a 5% subsample.
Intended use
Read the paper and explore the dataset and evaluation toolkit on GitHub ↗.