MiniCPM-SALA

MiniCPM Text Series

Use your own data to make MiniCPM-SALA fit your target task better. You can customize the training workflow with the Transformers Trainer, or follow the LLaMA-Factory guide to prepare data and configs for full-parameter or LoRA training.

If you want the model to better suit your business, persona or a specific task, you can train it with the fine-tuning tools we support.

Two approaches are currently recommended:

• LLaMA-Factory: better for beginners and general developers. Configuration is simple and it also provides a WebUI, making it suitable for quickly completing LoRA / QLoRA / SFT fine-tuning tasks.
• Transformers Trainer: better for developers with training experience, making it easy to customize data processing, training parameters and distributed training workflows.

You can choose a training approach based on your hardware:

Training environment	Suitable scenarios	Recommended approach
Single GPU	Quick validation, small datasets, persona fine-tuning, LoRA / QLoRA	LLaMA-Factory
Single node, multiple GPUs	Larger datasets, higher training efficiency	LLaMA-Factory / Transformers Trainer
Multi-node training	Large-scale training, research or production-grade training tasks	Transformers Trainer + a distributed training solution

If this is your first time fine-tuning, we recommend starting with a single GPU + LLaMA-Factory + LoRA / QLoRA. Once you've confirmed the data format, training workflow and model results, consider multi-GPU or distributed training.

---

Hugging Face Transformers Trainer

We provide official scripts that make it easy to adapt the pretrained MiniCPM-SALA when fine-tuning on downstream tasks. Our fine-tuning scripts use transformers.Trainer and DeepSpeed by default. Both full-parameter fine-tuning and LoRA fine-tuning are supported.

• Full-parameter fine-tuning: full-parameter fine-tuning updates all of the LLM's parameters throughout training.
• LoRA fine-tuning: LoRA updates only a small subset of parameters, enabling lightweight model tuning. We provide a LoRA implementation based on peft.

Prerequisites

1. Installation

Install all packages.

pip install -r requirements.txt

📎 Attachment: requirements.txt

1. Multi-turn conversation format

The multi-turn conversation fine-tuning example follows a conversation format convention that assigns different loss_mask values to different roles, so the loss for multiple replies can be computed in a single forward pass.

Example dataset format:

JSON

[
  {
    "messages": [
      { "role": "system", "content": "<系统提示文本>" },
      { "role": "user", "content": "<用户提示文本>" },
      { "role": "assistant", "content": "<助手回复文本>" },
      { "role": "user", "content": "<用户提示文本>" },
      { "role": "assistant", "content": "<助手回复文本>" }
    ]
  }
]

Note: The fine-tuning code now includes a validation set. A complete dataset must contain a training set and a validation set (a test set is optional). The data format of a single example is as follows:

JSON

{
  "messages": [
    {
      "role": "user",
      "content": "类型#裙*裙长#半身裙"
    },
    {
      "role": "assistant",
      "content": "这款百搭时尚的仙女半身裙，整体设计非常的飘逸随性，穿上之后每个女孩子都能瞬间变成小仙女啦。料子非常的轻盈，透气性也很好，穿到夏天也很舒适。"
    }
  ]
}

DeepSpeed ZeRO

Full-parameter Fine-tuning

Bash

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

deepspeed --include localhost:0,1,2,3,4,5,6,7 finetune.py \
    --model_name_or_path openbmb/MiniCPM-SALA \
    --output_dir output/AdvertiseGenSFT/$formatted_time/ \
    --train_data_path data/AdvertiseGenChatML/train.json \
    --eval_data_path data/AdvertiseGenChatML/dev.json \
    --learning_rate 5e-5 \
    --per_device_train_batch_size 2 \
    --per_device_eval_batch_size 32 \
    --bf16 \
    --gradient_accumulation_steps 4 \
    --warmup_steps 100 \
    --max_steps 3000 \
    --weight_decay 0.01 \
    --eval_steps 100 \
    --save_strategy steps \
    --save_steps 500 \
    --seed 42 \
    --log_level info \
    --logging_strategy steps \
    --logging_steps 10 \
    --deepspeed configs/ds_config_zero3.json

LoRA Fine-tuning

Simply add --use_lora to the script above.

Accelerate

Full-parameter Fine-tuning

Bash

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

accelerate launch --config_file configs/accelerate/fsdp_config.yaml \
    finetune.py \
    --model_name_or_path openbmb/MiniCPM-SALA \
    --output_dir output/AdvertiseGenSFT/$formatted_time/ \
    --train_data_path data/AdvertiseGenChatML/train.json \
    --eval_data_path data/AdvertiseGenChatML/dev.json \
    --learning_rate 5e-5 \
    --per_device_train_batch_size 2 \
    --per_device_eval_batch_size 32 \
    --bf16 \
    --gradient_accumulation_steps 4 \
    --warmup_steps 100 \
    --max_steps 3000 \
    --weight_decay 0.01 \
    --eval_steps 100 \
    --save_strategy steps \
    --save_steps 500 \
    --seed 42 \
    --log_level info \
    --logging_strategy steps \
    --logging_steps 10

LoRA Fine-tuning

Simply add --use_lora to the script above.

---

LLaMA-Factory

Our Hugging Face checkpoint openbmb/MiniCPM-SALA can be used directly in LLaMA-Factory for full-parameter and LoRA training.

We provide training guides for SFT, KTO, DPO and continued pretraining.

Prerequisites

1. Installation

Clone and install the LLaMA-Factory GitHub repository.

Bash

git clone https://github.com/hiyouga/LlamaFactory.git
cd LLaMA-Factory
pip install -r requirements.txt

1. Data preparation

Since we support four training modes (pretrain, sft, dpo, kto), you need to prepare your data following the format provided in {mode}_demo.json. Then add your dataset information to LLaMA-Factory/data/dataset_info.json so that your dataset can be found. Example:

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,
    "formatting": "sharegpt",
    "columns": {
      "messages": "conversations",
      "chosen": "chosen",
      "rejected": "rejected"
    }
  }
}

Creating the Training Config YAML File

Full-parameter Fine-tuning

Create a full training config file named minicpm_sala_sft.yaml and place it in LLaMA-Factory/examples/minicpm_config/. For other configs, see llama_factory/configs.

YAML

### model
model_name_or_path: openbmb/MiniCPM-SALA
trust_remote_code: true

### method
stage: sft
do_train: true
finetuning_type: full

### ddp
ddp_timeout: 180000000
deepspeed: examples/deepspeed/ds_z3_config.json

### dataset
dataset: sft_zh_demo
template: cpm4
cutoff_len: 1800
max_samples: 500000
overwrite_cache: true
preprocessing_num_workers: 16

### output
output_dir: saves/minicpm/minicpm_sala_full
logging_steps: 10
save_strategy: epoch
plot_loss: true
overwrite_output_dir: true

### train
per_device_train_batch_size: 2
gradient_accumulation_steps: 4
learning_rate: 0.0001
num_train_epochs: 3.0
lr_scheduler_type: cosine
warmup_ratio: 0.1
bf16: true

### eval
val_size: 0.1
per_device_eval_batch_size: 4
eval_steps: 500

LoRA Fine-tuning

Create a config file named minicpm_sala_lora_sft.yaml and place it in LLaMA-Factory/examples/minicpm_config/. Copy the contents of minicpm_sala_sft.yaml above and simply change finetuning_type from full to lora.

Model Training Commands

Full Training

1. DeepSpeed

Bash

llamafactory-cli train examples/minicpm_config/minicpm_sala_sft.yaml

1. Accelerate FSDP multi-node

Bash

accelerate launch \
    --config_file examples/accelerate/fsdp_config_multiple_nodes.yaml \
    src/train.py examples/minicpm_config/minicpm_sala_sft.yaml

LoRA Training (DeepSpeed)

Bash

llamafactory-cli train examples/minicpm_config/minicpm_sala_lora_sft.yaml