mesolitica/MalaysianFeedback

Malaysian Feedback

Well captured diversity preference dataset to train reward models to make the models more Malaysian.

We collect from Twitter, Facebook, c.cari.com.my forum, b.cari.com.my forum and Lowyat forum to make sure the dataset is Malaysian diverse enough.

Dataset construction

Instruction Sampling

We parsed questions from different social media, and use these parsed questions as the instruction dataset.

Model Sampling

We select different base models of all levels guided by MaLLaM as knowledge steering to complete the instructions. We use,

1. GPT-4o
2. GPT-4o Mini
3. Sonnet 3.5 V2
4. Haiku 3.5
5. Llama 3.1 70B Instruct
6. Qwen2.5 72B Instruct