Help Center/ ModelArts/ Best Practices/ LLM Training/ Training the Qwen3 Series Models Adapted to the MindSpeed-LLM Framework on ModelArts/ Preparing Model Weights and Training Data
Updated on 2026-09-03 GMT+08:00
View PDF
Share

Preparing Model Weights and Training Data

Obtaining the Model Weights and Code

This section uses Qwen models as an example to describe the training process. Table 1 lists the supported LLMs and the addresses for obtaining their weights.

Table 1 Supported LLMs and model weight sources

Model

Training Scenario

Training Framework

Open-Source Weight File Link

Qwen3-8B

SFT

MindSpeed-LLM

https://huggingface.co/Qwen/Qwen3-8B

Qwen3-30B-A3B

SFT

MindSpeed-LLM

https://huggingface.co/Qwen/Qwen3-30B-A3B

Qwen3-32B

SFT

MindSpeed-LLM

https://huggingface.co/Qwen/Qwen3-32B

To download weight files from the Hugging Face website, set up a proxy. For instructions, search online.

Upload the downloaded model weight files to your OBS bucket. The following is an example of the directory for storing files in your OBS bucket based on the OBS plan:

obs://mindspeed-llm/mindspeed-llm-a2/Qwen3-8B
obs://mindspeed-llm/mindspeed-llm-a2/Qwen3-32B
obs://mindspeed-llm/mindspeed-llm-a2/Qwen3-30B-A3B

Obtaining Training Data

The Qwen3 series models are used as an example. The following dataset can be used:

  • Fine-tuning dataset

    alpaca_gpt4_data is used as an example.

    wget https://huggingface.co/datasets/QingyiSi/Alpaca-CoT/blob/main/alpacaGPT4/alpaca_gpt4_data.json