Large Language Model Alignment via Direct Preference Optimization for Indonesian Technical Documentation
Views: 3
Downloads: 0
Sri Kurniawan
1
,
Gita Pratama
2
,
Dewi Kurniawan
3
1, 2, 3Department of Forestry and Environmental Science, IPB University
Downloads
Reinforcement Learning from Human Feedback (RLHF) incurs substantial reward model instability during specialized domain alignment. Applying Direct Preference Optimization directly on paired engineering responses, our fine-tuned 13B model achieves state-of-the-art accuracy in Indonesian software technical support synthesis.
