Appearance
微调
教程
- 通俗易懂理解全量微调和 LoRA 微调 从线性代数理解 LoRA 及其 rank
- 如何微调大模型并暴露接口给后端调用
- 代码仓库/数据集:https://github.com/huangyf2013320506/magic_conch_backend
- 看点:在 AutoDL 上使用 LLaMA-Factory 进行可视化微调
- 如何把你的 DeePseek-R1 微调为某个领域的专家?
- 配套资料: 《人人都能学会的 AI 教程》资源合集
- 看点:使用 硅基流动 微调大模型 jsonl 数据
- Use Ollama with any GGUF Model on Hugging Face Hub 本地运行 HuggingFace 的模型
- 13 分钟一起微调一个开源大模型:用 LoRA 让模型掌握新知识
平台和工具
平台
- colab
- http://kaggle.com/
- https://huggingface.co/
- https://hf-mirror.com/
- https://autodl.com/
- https://cloud.siliconflow.cn/
工具
- 客户端
- chatbox
- anythingLLM
- 微调框架
- unsloth
- llamafactory.cn
- 本地运行
- ollama:gguf
- vllm:HF 模型目录(safetensors)、OpenAI API
- 其它
- modelscope
- evalscope
- huggingface-cli
草稿
- 多少数据集合适
- 社区镜像 agiclass/fine-tuning-lab/finetune-lab-v8:v1
- 连接 ssh
- 连接 sftp
- scp
- 专门 SFTP 客户端
- hf-mirror 下载预模型 HuggingFace
- tmux nohup
- colab 微调大模型 unsloth.ai
- 将微调后的大模型推送到 huggingface
- 使用微调后的大模型
- 本地使用 ollama 运行大模型
混合推理 对话+思考
主流微调工具:
高层:unsloth(单机单卡)、Llama-Factory
底层:peft、transformer
必须要会使用 vscode 远程服务器
# 你的 gradio 和 transformers 要求 huggingface_hub < 1.0。那我们装:
pip install -U "huggingface_hub<1.0"
huggingface-cli download --resume-download deepseek-ai/DeepSeek-R1-Distill-
Qwen-1.5B