Skip to main content
Back

MiniCPM4.1 & MiniCPM4

MiniCPM Text Series

Choose between the official code and the LLaMA-Factory route based on your training goal, and learn about data preparation, training configuration, and methods such as LoRA / QLoRA. Resource requirements differ across models and training configurations, so check the target version, your hardware and the sample parameters first.

Hardware requirements​

Model sizeMinimum GPUs for full fine-tuningMinimum GPUs for LoRA fine-tuning
0.5BAt least one 20-series GPU with more than 12GB of VRAMAt least one 20-series GPU with 8GB of VRAM
8BAt least four GPUs with 24GB of VRAMAt least two 20-series GPUs with 24GB of VRAM

Device requirements: at least one GPU with 24GB of VRAM, 20-series or newer; for QLoRA you can try a 12GB GPU

Get the official code with git​

Bash
git clone https://github.com/OpenBMB/MiniCPM

Prepare the dataset​

For dialogue datasets, convert them into the following JSON format, where each entry is one messages list

The system field highlighted in red below is optional:

JSON
[
  {
    "messages": [
      {
        "role": "system",
        "content": "<system prompt text>"
      },
      {
        "role": "user",
        "content": "<user prompt text>"
      },
      {
        "role": "assistant",
        "content": "<assistant response text>"
      },
      // ... Muti Turn，多轮对话
      {
        "role": "user",
        "content": "<user prompt text>"
      },
      {
        "role": "assistant",
        "content": "<assistant response text>"
      }
    ]
  }
]

Edit the MiniCPM/finetune/lora_finetune_ocnli.sh file​

Bash
formatted_time=$(date +"%Y%m%d%H%M%S")
echo $formatted_time

#4090添加这两行代码
export NCCL_P2P_DISABLE=1
export NCCL_IB_DISABLE=1 

deepspeed --include localhost:1 --master_port 19888 finetune.py \
    --model_name_or_path openbmb/MiniCPM3-4B \ # 可以修改为本地模型目录
    --output_dir output/OCNLILoRA/$formatted_time/ \ # 可以修改为其他用来保存输出模型的地址
    --train_data_path data/ocnli_public_chatml/train.json \ # 这里写按照第二步处理好的训练集地址
    --eval_data_path data/ocnli_public_chatml/dev.json \ # 这里写按照第二步处理好的验证集
    --learning_rate 5e-5 \ # 学习率
    --per_device_train_batch_size 16 \ # 每张卡训练时的batch_size
    --per_device_eval_batch_size 128 \ # 每张卡测试时的batch_size
    --model_max_length 1024 \ # 模型训练时最大token数，超出将截断
    --bf16 \ # 是否使用bf16数据格式，如果不是改为false
    --use_lora \ # 是否使用lora
    --gradient_accumulation_steps 1 \ # 梯度累计次数
    --warmup_steps 100 \ # 预热步数
    --max_steps 1000 \ # 最大训练步数，到达后停止训练
    --weight_decay 0.01 \ # 权重正则化值
    --evaluation_strategy steps \ # 测试方法，可以改为epoch
    --eval_steps 500 \ # 与evaluation_strategy steps一起起作用，500个step测试一一次
    --save_strategy steps \ # 模型保存策略，可以改为epoch即每个epoch保存一次
    --save_steps 500 \ # save_strategy steps一起起作用，代表500步保存一次
    --seed 42 \ # 随机种子
    --log_level info --logging_strategy steps --logging_steps 10 \ # logging的设置
    --deepspeed configs/ds_config_zero2_offload.json # deepspeed配置文件设置，如果显存充足可以改为configs/ds_config_zero2_offload.json

Train with LoRA / QLoRA (optional)​

lora​

Add the use_lora parameter to the MiniCPM/finetune/lora_finetune_ocnli.sh file.

If you're unsure where, add the following line just before the --deepspeed configs/ds_config_zero2_offload.json line:

Bash
--use_lora \

qlora​

Add the use_lora and qlora parameters to the MiniCPM/finetune/lora_finetune_ocnli.sh file.

If you're unsure where, add the following lines just before the --deepspeed configs/ds_config_zero2_offload.json line:

Bash
--use_lora \
--qlora \

After editing the bash file above, run the following commands to start training:​

Bash
cd MiniCPM/finetune
bash lora_finetune_ocnli.sh

Usage is essentially the same as in [Fine-tuning Tutorial] MiniCPM2.0. If you are not training tool-calling capability, you only need to change template: cpm3. For training tool-calling capability, stay tuned for the PR we are submitting to llama-factory.

First, install the Llama_Factory dependencies​

Bash
git clone https://github.com/hiyouga/LLaMA-Factory.git
cd LLaMA-Factory
pip install -e ".[torch,metrics]"

Process the data​

Convert your dataset into the format used in the Minicpm/finetune/llama_factory_example/llama_factory_data folder (the examples cover three fine-tuning methods: DPO, KTO and SFT), and place it in the llama_factory/data directory.

DPO data format​

Below is an example of DPO data for constructing multi-turn dialogues in Llama_Factory, where:

JSON
{
    "instruction": "你是一个科技百科小助手，能够回复大部分科技领域问题",
    "input": "我想知道 MiniCPM4 到底好不好呀？",
    "chosen": "亲呀，MiniCPM4 真的很不错呢！它运用了 NSA 注意力机制呀，这使得它在处理信息等方面有着很出色的表现，能够更加精准地理解和生成内容，无论是文本创作还是辅助学习等场景，都可以发挥挺不错的作用呢，很多用户使用后都反馈体验挺好的哦。",
    "rejected": "哎呀，MiniCPM4 嘛，就是一般般啦。它虽然有一些功能，但是在实际使用中，也没有特别突出的亮点呀，和其他同类产品比起来，优势也不是特别明显呢，反正就是能用，但也没达到让人惊艳的程度啦。",
    "history": [
        [
        "我听说 MiniCPM4 挺智能的，是真的吗？",
        "亲，那确实呀，MiniCPM4 挺智能的呢，它有着独特的技术加持，在很多任务上都能展现出不错的智能水平，能帮大家解决不少问题哦。"
        ],
        [
        "那它在语言理解方面表现如何呢？",
        "亲，它在语言理解方面的表现可圈可点哦，靠着相关的技术，像 NSA 注意力机制助力，对各种语句的理解还是挺准确的呢，能较好地把握语义呀。"
        ],
    ]
}

KTO data format​

  1. Each messages represents one data entry, and label indicates whether the reply in that entry is reasonable
  2. Unlike DPO, you don't need different answers to the same question; instead, give a rating to each question-answer pair
  3. Each question and reply can be rated True (good) or False (bad)
  4. The places you need to modify are highlighted in red
JSON
{
    "messages": [
      {
        "content": "minicpm4.0哪里最强？",
        "role": "user"
      },
      {
        "content": "哪里都强，六边形战士好吗",
        "role": "assistant"
      }
    ],
    "label": true
  },
  {
    "messages": [
      {
        "content": "minicpm3.0 有哪些优点。",
        "role": "user"
      },
      {
        "content": "1.长文本能力。2.function call能力 3.其他所有能力",
        "role": "assistant"
      },
      {
        "content": "细说长文本能力",
        "role": "user"
      },
      {
        "content": "反正很强，细说就去我们github吧。https://github.com/OpenBMB/MiniCPM.git",
        "role": "assistant"
      }
    ],
    "label": false
  }
  ]

The SFT data format is as follows; each dictionary forms one data entry​

alpaca format:

JSON
[
  {
    "instruction": "你现在比牛顿还牛的科学家",
    "input": "识别并解释给定列表中的两个科学理论：细胞理论和日心说。",
    "output": "细胞理论是生物科学的一个理论，它认为所有生命体都是由微小的基本单元——细胞所构成。这是生物学的一个基础理论，认为细胞是所有生物的基本结构和功能单位，所有的生物都是由一个或多个细胞组成，细胞只能通过细胞分裂产生新的细胞。这一理论由薛定谔、施瓦内和雪莱于1839年首次提出。\n\n日心说是指太阳是太阳系的中心，也就是说，行星围绕太阳旋转的理论。这个理论打破了传统的地心说观点，认为地球并不是宇宙的中心。日心说的提出者是尼古拉·哥白尼，他在16世纪初发表了他的著作《天体运行论》，阐述了太阳系行星围绕太阳运行的模型，为天文学的发展做出了巨大贡献。"
  }
 ]

chatml format:

JSON
[
  {
    "messages": [
      {
        "content": "Who are they?",
        "role": "user"
      },
      {
        "content": "They're Kane and Gretzka from Bayern Munich.",
        "role": "assistant"
      },
      {
        "content": "What are they doing?<image>",
        "role": "user"
      },
      {
        "content": "They are celebrating on the soccer field.",
        "role": "assistant"
      }
    ]
  }
]

Add the dataset information to dataset_info.json​

Add your dataset information to llama_factory/data/dataset_info.json so that your dataset can be found there, as in the example below:

  1. file_name is the file name of the dataset you created
  2. For example, "sft_zh_demo" is the dataset name, which will be used in the yaml configuration file later
  3. Only two things need to be changed: the key name of the dataset, e.g. "sft_zh_demo", and the data file name for file_name, e.g. alpaca_zh_demo.json
  4. All places that need to be modified are highlighted in red
JSON
{"identity": {
    "file_name": "identity.json"
  },
    "sft_zh_demo": {
      "file_name": "alpaca_zh_demo.json"
    },
    "kto_en_demo": {
      "file_name": "kto_en_demo.json",
      "formatting": "sharegpt",
      "columns": {
        "messages": "messages",
        "kto_tag": "label"
      },
      "tags": {
        "role_tag": "role",
        "content_tag": "content",
        "user_tag": "user",
        "assistant_tag": "assistant"
      }
    },
    "dpo_en_demo":{
    "file_name":"dpo_en_demo.json",
    "ranking":"true",
    "columns":{
      "prompt":"instruction",
      "query":"input",
      "chosen":"chosen",
      "rejected":"rejected",
      "history": "history"
    }
  }
}

Set up the training script​

Copy the files from MiniCPM/finetune/llama_factory_example into the LLaMA-Factory/examples/minicpm directory​

Bash
cd LLaMA-Factory/examples
mkdir minicpm
#以下代码中的/your/path要改成你的MiniCPM代码和LLaMA-Factory路径
cp -r /your/path/MiniCPM/finetune/llama_factory_example/*  /your/path/LLaMA-Factory/examples/minicpm

Depending on the fine-tuning method you need (DPO as an example), the configuration parameters in LLaMA-Factory/examples/minicpm/minicpm_dpo.yaml that must be modified are as follows​

YAML
model_name_or_path: openbmb/MiniCPM3-4B # 或者你本地保存的地址
template: cpm3 #注意如果微调minicpm3或者minicpm4就写cpm3，如果是minicpm2则写cpm
dataset: dpo_en_demo # 这里写dataset_info.json中的键名
output_dir: your/finetune_minicpm/save/path # 你微调后模型的保存地址
bf16: true # 如果你的设备支持bf16，否则false
deepspeed: examples/deepspeed/ds_z2_config.json # 如果显存不够可以改成ds_z3_config.json

If you need LoRA fine-tuning, change the following in the yaml file:​

Bash
finetuning_type: lora
lora_target: all

Start training​

Bash
cd LLaMA-Factory
FORCE_TORCHRUN=1 llamafactory-cli train examples/train_full/MiniCPM4.yaml

Developer Showcase

View all →
MiniCPM TextEdgeHome HarnessA replayable, evaluable local control loop for a 1B on-device model, built in RustMiniCPM TextSqueezing a NAS: Hooking Up WorkBuddyRunning local AI on an N150 CPU: real-world uses and lessons learnedMiniCPM TextWhen a Catgirl Learns to Think, on Your PhoneWas this model made for catgirls? An on-device deployment testMiniCPM TextMaking a 1B Model a Good Math AgentTeXada’s on-device MiniCPM practice and lessonsMiniCPM TextBMB Agent: a MiniCPM-powered Personal AssistantLetting a 2B on-device model take over schedules and todos, with real-world lessonsMiniCPM TextSelf-Evolving Catgirl: Reasoning via On-Policy Self DistillationNo teacher, no new base model: math reasoning emerges from self-distillation

On this page

Developer ShowcaseProjects from community developersCommunity EventsJoin the OpenBMB open-source community