Fine‑Tune DeepSeek‑R1 with MaxCompute & DataWorks on Alibaba Cloud

Overall Overview

Based on Alibaba Cloud's cloud‑native big data service MaxCompute and the data governance platform DataWorks, this article demonstrates how to fine‑tune the DeepSeek‑R1 distilled model using a custom dataset. The workflow consists of two main parts: fine‑tuning on the AI platform PAI and constructing & connecting a custom dataset.

How to Fine‑Tune DeepSeek

In the PAI console, go to Quick Start > Model Gallery and select a model, e.g., DeepSeek‑R1‑Distill‑Qwen‑7B. Click the “Train” button; the core step is to fine‑tune the model with a custom dataset, which can be stored in OSS or MaxCompute. The demo uses a dataset stored in MaxCompute.

After selecting the model page, click “Train”. Choose a custom dataset, create a new dataset, and set the storage type to MaxCompute.

Import the MaxCompute project and table names, configure the path, and then select the model output path.

Choose the resource and parameter configuration, then click “Train” to start fine‑tuning with your custom dataset.

How to Build a Custom Dataset and Connect It to DeepSeek

In PAI, creating a custom dataset requires linking a MaxCompute project and table. First, create a MaxCompute project via the MaxCompute console (Workspace → Project Management → New Project).

Then create a MaxCompute table through the DataWorks console (Data Development → Table Management → New Table). Detailed steps are documented in the linked guide.

Writing Data to the Custom Dataset

After publishing the table schema, you can write data using DataWorks integration tasks or MaxCompute node tasks. DataWorks also supports uploading local files directly to MaxCompute tables.

Deploying and Using the Fine‑Tuned DeepSeek‑R1 Model

Once the project and tables are set up, use the PAI platform to fine‑tune the DeepSeek‑R1 distilled model with your custom dataset. Additional tutorials for deploying DeepSeek‑V3, DeepSeek‑R1, and other large language models are available via the provided links.

For a quick start, you can also deploy the Tongyi Qianwen model with one click on PAI or deploy multimodal MLLM applications via EAS.

Developers are invited to scan the QR code below to join the user community for further assistance.