qwen2.5-3B-instarl微调

全参数微调

阿里云人工智能平台API/交互式建模DSW

1.微调准备（网页UI安装）

1.1安装modelscope

要求PyTorch 版本 ≥ 2.6.0

pip install modelscope

pip install addict

1.2下载安装ollama

modelscope download --model gfunsAi/ollama-linux --local_dir ./ollama-linux --revision v0.6.0

cd ollama-linux && sudo chmod 777 ./ollama-modelscope-install.sh && ./ollama-modelscope-install.sh

1.3下载安装LLaMA-Factory

git clone https://github.com/hiyouga/LLaMA-Factory.git

cd LLaMA-Factory

pip install -e .[metrics]

验证安装是否成功：

llamafactory-cil version

1.4下载qwen2.5-3B-Instruct

from modelscope.hub.snapshot_download import snapshot_download

model_id = "Qwen/Qwen2.5-3B-Instruct"
local_dir = "./models/Qwen/Qwen2.5-3B-Instruct"

# 下载模型
model_dir = snapshot_download(model_id, local_dir=local_dir)
print(f"Model downloaded to: {model_dir}")

ollama-linux/LLaMA-Factory/cache

cd ollama-linux/LLaMA-Factory/

pip install --upgrade “transformers>=4.45.0”

DISABLE_VERSION_CHECK=1 llamafactory-cli webui --torch_dtype bfloat16

将.json文件送入到data文件夹中

聊天

训练

修改ollama-linux/LLaMA-Factory/data/dataset_info.json中为自己数据



saves/Qwen2.5-3B-Instruct/lora/train_2025-09-11-17-06-25

训练后预测

将检查点位置改为checkpoint位置

删除文件夹：

rm -r 文件夹名称