File size: 661 Bytes
602ad26 d7f7511 e1efacd d7f7511 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 |
---
language:
- ms
- en
---
# 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. Llama 3.1 70B Instruct
2. Qwen2.5 72B Instruct
|