Recent content on Chan Kha Vu 🇺🇦
I joined a team and we trained 7B and 14B math reasoning models based on DeepSeek-R1-Distill using SFT and GRPO. Our 14B model achieved 75.8% Maj@32 on AIME’25 ( +8.7% improvement), and our 7B model reached 65.8% Maj@32 ( +7.5% ). Here is what I’ve learned.
Turns out, the RL training process of DeepScaleR-1.5B introduced only low-rank changes to its base model, DeepSeek-R1-Distill-1.5B.
Learning to Rank (LTR) is a core component of any recommendation system. It is the algorithm that forms the final list of items to be shown to the user. This blog post is a comprehensive introduction to the basics of LTR and Unbiased LTR. Hopefully, it will give you enough context to build your own models or to understand more recent research in the field.
I built a cheap-ish 2x3090 RTX Deep Learning rig for my personal projects and experiments. In this post, I will share my build philosophy, inspirations, and reasoning behind my component choices.
To celebrate the week of Bing’s integration with ChatGPT, I built an AI bot based on GPT-3 and BLIP-2 to play Dixit and gathered some friends and co-workers to play against it.
A brief survey of common supervised approaches for Deep Metric Learning, as well as the new methods proposed in recent years.
Do you want to dominate your interviewer? Wanna bring him to his knees? Make him think that you’re superior? I’ll tell you how!
An in-depth survey on Multi-Task Learning techniques that works like a charm as-is right from the box and are easy to implement – just like instant noodle!
The connection between the proof of the existance of the Lorenz Attractor, its Geometric Flow Model, and the Knot Theory Model.
Hi, my name is Chan Kha Vu. I’m a ML engineer at Microsoft in Seattle Area, currently working on image ranking models at Bing . Every time you search for images on Bing , the final results page are formed by the models trained with my code! Previously, I worked at Samsung Research on Monocular 3D Reconstruction, Multi-Task Learning, and optimizing Conv Nets for mobile devices. I will never…