MY-FairnessBench

LLM benchmark · 2026 · Archive 04 / 11

A command-line benchmark for testing model bias in a Malaysian cultural, linguistic, and legal context.

Most fairness benchmarks flatten away the details that matter in Malaysia. MY-FairnessBench keeps ethnicity, socio-economic history, Manglish, and code-switching in the test instead.

What I made

The CLI sends the same prompt through a set of international models and ILMU, Malaysia's first multimodal LLM. It then saves the responses to a timestamped Markdown report for comparison.

This is still a research prototype. The useful part is the framing: local bias cannot be measured well with examples written for a different social and legal setting.

How it works

5 steps
01Write a case

Frame a prompt around Malaysian context.

02Run models

Send the same case through each configured provider.

03Collect

Keep the raw responses together.

04Compare

Look for differences in language and treatment.

05Save

Write a dated Markdown report.

A compact map of the main product loop.

Repository breakdown

This comes from GitHub's detected language breakdown. It measures source size, not the time or difficulty of the work.

Language mix

GitHub source bytes
TypeScript100%
Role

Independent research build

Status

Research prototype

Source

Private GitHub repository