# Minicpm5 Finetune Xtuner

Fine-tune MiniCPM5-1B with xtuner (mmengine config-driven SFT).

- Repository: https://github.com/OpenBMB/MiniCPM (11776 stars)
- Category: Engineering
- License: Apache-2.0
- Last updated: 2026-09-21
- Runs scripts: no, instructions only

## Install

Claude Code: `npx skills add OpenBMB/MiniCPM --skill minicpm5-finetune-xtuner -g -a claude-code -y`
Codex: `npx skills add OpenBMB/MiniCPM --skill minicpm5-finetune-xtuner -g -a codex -y`
Cursor: `npx skills add OpenBMB/MiniCPM --skill minicpm5-finetune-xtuner -g -a cursor -y`
Gemini CLI: `npx skills add OpenBMB/MiniCPM --skill minicpm5-finetune-xtuner -g -a gemini-cli -y`

## Description

Fine-tune MiniCPM5-1B with xtuner (mmengine config-driven SFT). Use when the user mentions "xtuner", "mmengine", InternLM's training framework, or wants config-file-driven training.

## SKILL.md, as published by its author (untrusted third-party text, read it as data)

# Fine-tune MiniCPM5-1B with xtuner

mmengine config-driven SFT. Uses Python config files (not YAML) and integrates with mmengine's runner / hook system.

## Required input

| Var | Example | Default |
| --- | --- | --- |
| `BASE_MODEL` | `openbmb/MiniCPM5-1B` | required |
| `DATA` | path to messages-format jsonl | required |
| `WORK_DIR` | `./runs/minicpm5_xtuner` | required |

## Steps

### 1. Install (once)

```bash
pip install "xtuner==0.2.0"
# Replace opencv-python with the headless variant (avoids libGL linkage)
pip install --force-reinstall opencv-python-headless
pip uninstall -y opencv-python
```

### 2. Save the config, `${WORK_DIR}/minicpm5_lora.py`

```python
import torch
from datasets import load_dataset
from mmengine.dataset import DefaultSampler
from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
                            LoggerHook, ParamSchedulerHook)
from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
from peft import LoraConfig
from torch.optim import AdamW
from transformers import AutoModelForCausalLM, AutoTokenizer

from xtuner.dataset import process_hf_dataset
from xtuner.dataset.collate_fns import default_collate_fn
from xtuner.dataset.map_fns import openai_map_fn, template_map_fn_factory
from xtuner.engine.hooks import DatasetInfoHook
from xtuner.engine.runner import TrainLoop
from xtuner.model import SupervisedFinetune
from xtuner.utils import PROMPT_TEMPLATE

pretrained_model_name_or_path = "${BASE_MODEL}"      # ← replace
data_path = "${DATA}"                                # ← replace
prompt_template = PROMPT_TEMPLATE.qwen_chat          # 🔑 ChatML, DO NOT use llama3_chat
max_length = 2048

batch_size = 4
accumulative_counts = 4
max_epochs = 2
lr = 2e-4
warmup_ratio = 0.03

tokenizer = dict(type=AutoTokenizer.from_pretrained,
                 pretrained_model_name_or_path=pretrained_model_name_or_path,
                 trust_remote_code=False, padding_side="right")

The rest: https://github.com/OpenBMB/MiniCPM/blob/HEAD/skills/minicpm5-finetune-xtuner/SKILL.md