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