The dataset is currently empty. Upload or create new data files. Then, you will be able to explore them in the Dataset Viewer.
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,
- Llama 3.1 70B Instruct
- Qwen2.5 72B Instruct
- Downloads last month
- 34