Model Download

WeClone uses Qwen2.5-7B-Instruct by default. Other models supported by LLaMA Factory also work (thinking models are not recommended).

TIP

7B models generally give weak results (some users report acceptable results, which may depend on the data), 14B is barely acceptable, and 32B is better.

Method 1: Hugging Face

hf download Qwen/Qwen2.5-7B-Instruct --local-dir models/Qwen2.5-7B-Instruct

Method 2: ModelScope (recommended in mainland China)

uv pip install modelscope
modelscope download --model Qwen/Qwen2.5-7B-Instruct --local_dir ./models/Qwen2.5-7B-Instruct

⚠️ The model files are large and the download may take a while. Make sure you have enough disk space (at least 20GB is recommended).
If the connection drops, run the command again to resume.

Using another model

If you chose another model or downloaded it to another directory, change settings.jsonc:

"common_args": {
    "model_name_or_path": "your_model_path",
    "template": "your_model_template",
    ...
},
"train_sft_args": {
    "lora_target": "module_names_to_fine_tune", // e.g. "q_proj,v_proj"
    ...
},

Multimodal models

To fine-tune a multimodal model, change settings.jsonc:

• set model_name_or_path to a vision-language model such as ./models/Qwen2.5-VL-7B-Instruct
• set template to qwen2_vl
• add "image" to include_type
• append ,visual.merger.mlp.0,visual.merger.mlp.2 to lora_target

English chats

For chats mainly in English, translate default_system and set test_model_args.test_data_path to dataset/eval/test_data-en.json.