MiniCPM-V MiniCPM-V & o Cookbook

MiniCPM-V 4.6 微调教程 (LlaMA-Factory)

模型与任务概览

本章节以 counting 任务 allenai/pixmo-count 作为微调示例。

训练任务:

环境说明

最小可运行安装步骤

conda create -n "MiniCPM-V-4.6-Counting" python=3.11 -y
conda activate "MiniCPM-V-4.6-Counting"

pip install torch==2.8.0 torchvision==0.23.0

pip install \
  transformers==5.7.0 accelerate==1.13.0 \
  deepspeed==0.18.3 peft==0.18.1 trl==0.24.0 \
  wandb ninja einops safetensors tokenizers sentencepiece

MAX_JOBS=32 NVCC_THREADS=4 pip install --no-build-isolation flash-attn==2.8.3
git clone https://github.com/hiyouga/LlamaFactory.git
cd LlamaFactory
pip install -e .
pip install -r requirements/metrics.txt -r requirements/deepspeed.txt
pip install transformers==5.7.0

注意:MiniCPM-V 4.6 在 transformers>=5.7.0 版本被官方支持。请保证最后安装的 transformers 版本为 5.7.0 及以上。

依赖版本参考

python                        3.11.0
accelerate                    1.13.0
deepspeed                     0.18.3
flash_attn                    2.8.3
llamafactory                  官方最新代码
torch                         2.8.0
torchvision                   0.23.0
transformers                  5.7.0

数据准备

allenai/pixmo-count 下载数据集,并将其转换为 json 格式。

数据格式参考: json { "messages": [ { "content": "<image>\nCarefully observe the image. Are there any people in the image? If yes, please list their respective coordinates and provide the total count. If no, answer 0.", "role": "user" }, { "content": "<think>\n\n</think>\n\nThe respective coordinates of people: <point>236 469</point>, <point>307 232</point>, <point>362 434</point>, <point>485 521</point>, <point>487 340</point>, <point>615 386</point>, <point>735 441</point>, <point>870 615</point>. So the total count is 8.", "role": "assistant" } ], "images": [ "/path/to/images/*.jpg" ], "source_file": "pixmo-count", "orig_index": 1, "channel": "pixmo-count" } - Counting 任务中加入对 points 预测的监督能够提高微调的效果,因此我们推荐将数据中的 points 坐标拼到 assistant 回复中。 - 由于 MiniCPM-V 4.6 会将图片坐标归一化到 0~1000,因此也需要对 points 坐标进行以下处理: python def expected_norm(x_px: float, y_px: float, width: int, height: int) -> Tuple[int, int]: return int((x_px / width) * 1000.0), int((y_px / height) * 1000.0) - 微调训练推荐在 assistant 前缀中加入 <think>\n\n</think>\n\n。(如果是 thinking 的任务,则加入 <think>\n

启动训练

配置好模型路径、训练集路径、验证集路径和输出目录后,执行以下脚本即可以开始训练。 - 配置 train.yaml

### model
model_name_or_path: /path/to/minicpm-v-4_6
trust_remote_code: true
flash_attn: fa2

### method
stage: sft
do_train: true
finetuning_type: full
freeze_vision_tower: false
deepspeed: LlamaFactory/examples/deepspeed/ds_z2_config.json

### dataset
dataset: pixmo_count_train
eval_dataset: pixmo_count_val
dataset_dir: /path/to/dataset_dir # dataset_dir should contain dataset_info.json file
template: minicpm_v_4_6
cutoff_len: 4096
preprocessing_num_workers: 16
dataloader_num_workers: 16
overwrite_cache: true

### output
output_dir: /path/to/output_dir
logging_steps: 1
save_steps: 132
save_total_limit: 30
eval_strategy: steps
eval_steps: 80
plot_loss: true
overwrite_output_dir: false
report_to: wandb

### train
per_device_train_batch_size: 1
per_device_eval_batch_size: 1
gradient_accumulation_steps: 16
learning_rate: 5.0e-6
num_train_epochs: 4.0
lr_scheduler_type: cosine
warmup_ratio: 0.05
bf16: true
max_grad_norm: 1000
ddp_timeout: 180000000
weight_decay: 0.1
adam_beta2: 0.95
#!/bin/bash
set -euo pipefail

export CUDA_VISIBLE_DEVICES="${CUDA_VISIBLE_DEVICES:-0,1,2,3,4,5,6,7}"
export NPROC_PER_NODE="${NPROC_PER_NODE:-8}"
export MASTER_PORT="${MASTER_PORT:-29632}"

export WANDB_API_KEY="${WANDB_API_KEY:-}"
export WANDB_PROJECT="${WANDB_PROJECT:-MiniCPMV46-Counting}"
export WANDB_RUN_NAME="${WANDB_RUN_NAME:-mcpmv46_count}"
export WANDB_NAME="${WANDB_NAME:-mcpmv46_count}"

# MiniCPMV 4.6 downsample mode: 4x for high-resolution, 16x for default
export DOWNSAMPLE_MODE="${DOWNSAMPLE_MODE:-4x}"

export DISABLE_VERSION_CHECK=1
# Activate the lfv46 conda environment

# IMPORTANT: Unset USE_V1 to use the v2 launcher
unset USE_V1

CONFIG_FILE="$(dirname "$0")/train.yaml"
OUTPUT_DIR="${OUTPUT_DIR:-/path/to/output_dir}"

echo "Training with config: $CONFIG_FILE"
echo "Output dir: $OUTPUT_DIR"

llamafactory-cli train "$CONFIG_FILE"

关键参数说明 - 训练支持 16x4x 两种视觉 Token 压缩率,通过 export DOWNSAMPLE_MODE="${DOWNSAMPLE_MODE:-4x}" 参数进行控制。 - 当前版本的 transformers 对 Qwen3.5 系列的 packing 训练的支持仍存在问题,目前请先不要使用 packing 模式,官方修复后本文档也会进行更新。

训练过程

https://wandb.ai/majy24-tsinghua-university/MiniCPMV46-Counting-LF/reports/Llama-Factory---VmlldzoxNjgyNzk4NQ

Llama-Factory 训练过程

评测结果

评测指标说明:

指标 说明
Acc@0 精确匹配率,即预测值与真实值完全一致
Acc@0 Top1 训练过程保存的所有 checkpoint 中,评测结果 Acc@0 的最高分数
Acc@0 Avg.Top3 训练过程保存的所有 checkpoint 中,评测结果 Acc@0 前三名的平均分数

下表展示了两种视觉 Token 压缩率设置下的评测结果:

模型 视觉 Token 压缩率 Acc@0 Top1 Acc@0 Avg.Top3
MiniCPM-V 4.6 16 46.5 N/A [1]
MiniCPM-V 4.6 4 51.8 N/A [1]
Fine-tuned model 16 78.4 78.1
Fine-tuned model 4 83.1 82.5

[1]: MiniCPM-V 4.6 为原始模型,未经微调,仅有一个 Acc@0 结果 (Acc@0 Top1),无法计算 Acc@0 Avg.Top3

输出样例:

```text Q: Carefully observe the image. Are there any airplanes in the image? If yes, please list their respective coordinates and provide the total count. If no, answer 0.

A: The respective coordinates of airplanes: 310 370, 360 275, 385 486, 402 180, 439 368, 474 611, 505 250, 532 448, 536 818, 597 328. So the total count is 10. ```

LlamaFactory sample