Achieve OpenAI o1-mini Level Reasoning with Open-Source Models

Author(s): Yu-Cheng Tsai

Originally published on Towards AI.

Performing Supervised Fine-Tuning (SFT) on DeepSeek R1’s Distilled Models with Your Domain Data

This member-only story is on us. Upgrade to access all of Medium.

Photo by Jorne Hermans on Unsplash

What Are DeepSeek’s R1 and Its Distilled Models?

FP8 is only available and introduced for Nvidia Hopper Series. The image is from Nvidia Hopper Series.

DeepSeek has released a big reasoning model (671B, 37B Activated parameters, MoE architecture), DeepSeek-R1, comparable to OpenAI’s o1. However, DeepSeek-R1 was trained and released in FP8 mixed precision, optimized for NVIDIA’s Hopper-series GPUs as shown above. If you don’t have access to these GPUs, converting the model from FP8 to other precision for use on the other GPUs (e.g. A100s) can be cumbersome. Alternatively, you can use vLLM for inference. This thread provides guidance on using DeepSeek’s R1 model. Please note, it is not lightweight! Fortunately, along with DeepSeek R1, a couple of distilled models are released on HuggingFace. Think of the distillation process as teaching: a larger, more complex model (the teacher) passes its knowledge to a smaller, more efficient model (the student). In this case, DeepSeek-R1 is the teacher, known for its advanced reasoning skills. The student models are supervised fine-tuned (SFT) using data generated by DeepSeek-R1, enabling them to mimic the teacher’s reasoning patterns.

Why Use Distilled Models?

Enhanced Reasoning… Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI