Skip to main content
Qwen Cloud home page
Docs

Fine-tuning overview

Improve model performance for your use case by fine-tuning Qwen models on Qwen Cloud with SFT and LoRA.

Fine-tuning adapts a pre-trained model to your specific use case using your own data. Compared to prompt engineering alone, fine-tuning can:
  • Improve accuracy on domain-specific tasks
  • Reduce latency by replacing few-shot prompts with learned behavior
  • Align outputs with human preferences or brand voice
  • Replace larger models with fine-tuned smaller models at lower cost

Prerequisites

Before creating a fine-tuning job, you need:

Training method

Fine-tuning on Qwen Cloud uses SFT (Supervised Fine-Tuning) -- you train the model on input/output pairs to teach it desired responses. Training data is formatted as ChatML messages with user/assistant turns.

Training mode

All fine-tuning jobs use LoRA (Efficient Training), which trains low-rank adapter weights on top of the frozen base model. This approach is faster and lower cost than full parameter training.

Supported models

ModelTraining methodTraining mode
Qwen3-14BSFTLoRA
You can also select a previously fine-tuned custom model as the base model to continue training iteratively. When creating a fine-tuning job, switch to the Custom Models tab in the base model selector to choose one of your own models.

Data format

Each line of your JSONL file contains a multi-turn conversation in ChatML format: 
{"messages": [
  {"role": "system", "content": "You are a helpful assistant."},
  {"role": "user", "content": "What is fine-tuning?"},
  {"role": "assistant", "content": "Fine-tuning adapts a pre-trained model to a specific task using custom training data."}
]}

Dataset tips

  • Scale: Fine-tuning requires at least several hundred high-quality examples. Thousands of examples typically yield better results.
  • Diversity: Cover the range of scenarios your model will encounter in production. Balance data proportions to match actual usage patterns.
  • Quality: Ensure assistant responses reflect the exact behavior you want. Low-quality outputs in training data lead to low-quality model behavior.

Next steps