AI Wednesdays: The One About Fine-Tuning (Round 2)

​Join us as we go for a second round into the evolving landscape of language models and fine-tuning approaches.

​More about the Sharings

• ​Isaac Lim (Data Scientist, GovTech) will share more on AI Practice's recent work on performing safety alignment on SEA-Lion-v2.1-Instruct for Singlish toxicity. The sharing will cover both technical and practical concepts around safety alignment, including preference alignment algorithms and scalable serverless training, along with some interesting experimental findings. (Technical Level: 300)

• ​Yoeven D Khemlani (CEO & Founder, JigsawStack) will showcase some of the models that JigsawStack has trained in-house and talk about how small models with fine tuning and specialisation can do great. (Technical Level: 200 - 300)
  

• ​James Chua (Researcher, Truthful AI) will share his paper on whether new reasoning models like DeepSeek, Gemini 2.0 and Qwen 2.5 actually have reasoning that faithfully explains how they "think". The talk will examine the new paradigm of model "thoughts" and how this improves performance in tests for faithful Chain-of-Though. Additionally, he'll discuss what incentivizes faithful explanation in model reinforcement learning. (Technical Level: 300)

​More about the Speakers

• ​Isaac is a Data Scientist from GovTech’s AI Practice. He mainly works on RAG, fine-tuning and LLM safety. He was previously an ML Engineer at a Seattle-based fintech startup working on municipal bond pricing.
  

• ​Yoeven D Khemlani is the CEO and Founder of JigsawStack, a company that specialises in developing developer-friendly AI SDK that automates backend tasks like web scraping, OCR, and translation with fast, specialized models, reducing maintenance and scaling pain for engineers.
  

• ​James Chua is a researcher in Truthful AI, an AI Safety organisation based in Berkeley. He also spends time in the Singapore AI Safety Hub. His current focus is evaluating truthfulness, situational awareness and reasoning in language models. In a past life, he was a machine learning engineer in industry.