Community Project Profile
OAT
Research-friendly framework for LLM online alignment
- Organisation
- Sea AI Lab (SAIL)
- Group
- International corporate lab
- Category
- LLM online alignment framework
- Status
- Actively maintained
- Started
- 2024-10
- Language / Form
- Python
- License
- Apache-2.0
- GitHub Stars
- 652
- Updated
- 2026-05-04
OAT is Sea AI Lab’s LLM alignment-training framework, aimed at post-training workflows such as reinforcement learning and preference learning.
What It Is
OAT stands for Online Alignment Training. It packages common LLM post-training workflows such as reinforcement learning, preference optimization, online sampling, and evaluation into a research-friendly framework.
It is not an end-user product, but a tool for model research and training teams.
AI Relevance
Model capability increasingly depends on post-training. Pretraining provides base knowledge; SFT, RLHF, DPO, online reinforcement learning, and related workflows determine whether a model is useful, stable, and aligned.
OAT matters because it turns complex alignment experiments into reusable engineering infrastructure.
Singapore Relevance
OAT shows that Sea AI Lab is not only building regional language models, but also model-training tooling. That matters for a Singapore homegrown tech company participating in foundation-model competition.
Future tracking should watch whether it is used in Sailor or other SAIL model-training pipelines.
Milestones
- 2024-10OAT repository created