How to fine-tune Qwen1.5 (LoRA, QLoRA)

This guide covers adapting the Qwen1.5 family. Memory per method and GPU counts appear below, and the training cost calculator estimates runtime. For method background see fine-tuning, LoRA and QLoRA.

Setup

• Use transformers>=4.37.0, per the Qwen1.5-7B-Chat card, to avoid KeyError: 'qwen2'.
• No trust_remote_code is needed, according to the card.
• Format training data with the tokenizer's chat template (apply_chat_template) so it matches serving.
• The card does not document LoRA target modules, so use your trainer's defaults.
• Training context beyond 32K is not covered by the card, which states 32K for all sizes.

License

The Chat card lists the tongyi-qianwen license. The LICENSE text for that checkpoint says:

• Commercial use with more than 100 million monthly active users requires requesting a license from Alibaba Cloud.
• You may not use the Materials or any output to improve any other large language model (excluding Tongyi Qianwen or derivative works thereof).
• Recipients must receive a copy of the agreement, modified files must carry notices of changes, and distributed copies must retain the line "Tongyi Qianwen is licensed under the Tongyi Qianwen LICENSE AGREEMENT, Copyright (c) Alibaba Cloud. All Rights Reserved."

Fine-tuned weights are derivatives, so these terms travel with your adapter or merged model. Check the license field on the exact size you adapt, since other sizes may differ.