多模态大模型微调框架 LlamaFactory 使用指南

配置文件是 yaml 格式的文件，命名也很清晰，包括训练参数，任务配置参数。

在训练上，官方给了很多示例文件，比如全量训练、lora 微调、qlora 微调等方法。

训练

uv run llamafactory-cli train examples/train_lora/qwen3_lora_sft.yaml 

# examples/train_lora/llama3_lora_sft.yaml
model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
stage: sft
do_train: true
finetuning_type: lora
lora_target: all
dataset: identity,alpaca_en_demo
template: llama3
cutoff_len: 1024
max_samples: 1000
overwrite_cache: true
preprocessing_num_workers: 16
output_dir: saves/llama3-8b/lora/sft
logging_steps: 10
save_steps: 500
plot_loss: true
overwrite_output_dir: true
per_device_train_batch_size: 1
gradient_accumulation_steps: 8
learning_rate: 1.0e-4
num_train_epochs: 3.0
lr_scheduler_type: cosine
warmup_ratio: 0.1
bf16: true
ddp_timeout: 180000000
val_size: 0.1
per_device_eval_batch_size: 1
eval_strategy: steps
eval_steps: 500

导出

llamafactory-cli export merge_config.yaml 

# examples/merge_lora/llama3_lora_sft.yaml
### model
model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
adapter_name_or_path: saves/llama3-8b/lora/sft
template: llama3
finetuning_type: lora
### export
export_dir: models/llama3_lora_sft
export_size: 2
export_device: cpu
export_legacy_format: false

推理

llamafactory-cli chat inference_config.yaml 

# examples/inference/llama3.yaml
model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
template: llama3
infer_backend: huggingface # choices: [huggingface, vllm]

评估

llamafactory-cli eval examples/train_lora/llama3_lora_eval.yaml 

# examples/train_lora/llama3_lora_eval.yaml
### model
model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
adapter_name_or_path: saves/llama3-8b/lora/sft
# 可选项
### method
finetuning_type: lora
### dataset
task: mmlu_test # mmlu_test, ceval_validation, cmmlu_test
template: fewshot
lang: en
n_shot: 5
### output
save_dir: saves/llama3-8b/lora/eval
### eval
batch_size: 4

微调 Qwen3 VL

模型准备

llamafactory-cli 可以自动下载模型，但是国内有时会超时，建议使用国内镜像网站。在命令行中执行如下：

export HF_ENDPOINT="https://hf-mirror.com"

选择一个指定的模型 Qwen/Qwen3-VL-2B-Instruct。

数据准备

llamafactory 中数据集的配置集中在 data/dataset_info.json。

{
"identity": {"file_name": "identity.json"},
"alpaca_en_demo": {"file_name": "alpaca_en_demo.json"},
"alpaca_zh_demo": {"file_name": "alpaca_zh_demo.json"},
"glaive_toolcall_en_demo": {
"file_name": "glaive_toolcall_en_demo.json",
"formatting": "sharegpt",
"columns": {"messages": "conversations", "tools": "tools"}
}
}

dataset_info.json 中 json 格式的文件，配置了需要使用的数据集。key 是数据集的名字，value 是具体参数。

例如：

• 数据集名称：alpaca_en_demo
• 数据集路径：alpaca_en_demo.json

具体数据集的格式，llamafactory 目前支持 alpaca 和 sharegpt 两种数据格式。

alpaca：

[{"instruction":"Describe a process of making crepes.","input":"","output":"Making crepes is an easy and delicious process! Enjoy!"},{"instruction":"Transform the following sentence using a synonym: The car sped quickly.","input":"","output":"The car accelerated rapidly."}]

sharegpt：

[{"messages":[{"content":"<audio>What's that sound?","role":"user"},{"content":"It is the sound of glass shattering.","role":"assistant"}],"audios":["mllm_demo_data/1.mp3"]}]

本次使用 coco-2014-caption，属于 sharegpt 格式，所以使用 sharegpt 格式来准备。

dataset_info.json 注册 coco 数据集的配置项

{
"coco-400": {
"file_name": "coco-400.json",
"formatting": "sharegpt",
"columns": {"messages": "conversations", "id": "id"},
"tags": {"role_tag": "from", "content_tag": "value", "user_tag": "user", "assistant_tag": "assistant"}
}
}

coco 数据集的格式如下：

配置参数

微调就选择 qlora 的方式，根据工程给的示例文件去修改，选择的示例文件是：

### model
model_name_or_path: Qwen/Qwen3-4B-Instruct-2507
quantization_bit: 4 # choices: [8 (bnb/hqq/eetq), 4 (bnb/hqq), 3 (hqq), 2 (hqq)]
quantization_method: bnb # choices: [bnb, hqq, eetq]
trust_remote_code: true
### method
stage: sft
do_train: true
finetuning_type: lora
lora_rank: 8
lora_target: all
### dataset
dataset: identity,alpaca_en_demo
template: qwen3_nothink
cutoff_len: 2048
max_samples: 1000
preprocessing_num_workers: 16
dataloader_num_workers: 4
### output
output_dir: saves/qwen3-4b/lora/sft
logging_steps: 10
save_steps: 500
plot_loss: true
overwrite_output_dir: true
save_only_model: false
report_to: none
### train
per_device_train_batch_size: 1
gradient_accumulation_steps: 8
learning_rate: 1.0e-4
num_train_epochs: 3.0
lr_scheduler_type: cosine
warmup_ratio: 0.1
bf16: true
ddp_timeout: 180000000
### eval
# val_size: 0.1
# per_device_eval_batch_size: 1
# eval_strategy: steps
# eval_steps: 500

根据以上模板，修改成我们自身的参数，关键修改在于：

1. 模型名称：model_name_or_path
2. 数据集：dataset
3. 模板：template 视觉大模型是 qwen3_vl_nothink，语言大模型是 qwen3_nothink，用错模板会报错

剩下的如训练批次、batch_size、梯度累计、学习率、保存路径、训练记录等都有设置，不再详说。

### model
model_name_or_path: Qwen/Qwen3-VL-2B-Instruct
quantization_bit: 4
quantization_method: bnb
trust_remote_code: true
### method
stage: sft
do_train: true
finetuning_type: lora
lora_rank: 8
lora_target: all
### dataset
dataset: coco-3000
template: qwen3_vl_nothink
cutoff_len: 2048
preprocessing_num_workers: 16
dataloader_num_workers: 4
### output
output_dir: saves/qwen3-2b-coco-3000/lora/sft
logging_steps: 10
save_steps: 500
plot_loss: true
overwrite_output_dir: true
save_only_model: false
report_to: none
### train
per_device_train_batch_size: 2
gradient_accumulation_steps: 4
learning_rate: 1e-5
num_train_epochs: 2
lr_scheduler_type: cosine
warmup_ratio: 0.1
bf16: true
ddp_timeout: 180000000
### eval
# val_size: 0.1
# per_device_eval_batch_size: 1
# eval_strategy: steps
# eval_steps: 500

启动训练

uv run llamafactory-cli train examples/train_qlora/qwen3-coco.yaml 

➜ LlamaFactory git:(main) ✗ uv run llamafactory-cli train examples/train_qlora/qwen3-coco.yaml
[WARNING|2026-02-06 17:47:42] llamafactory.hparams.parser:148>> We recommend enable `upcast_layernorm` in quantized training.
Qwen3VLVideoProcessor {"crop_size": null,"data_format":"channels_first","default_to_square": true,"device": null,"do_center_crop": null,"do_convert_rgb": true,"do_normalize": true,"do_rescale": true,"do_resize": true,"do_sample_frames": true,"fps":2,"image_mean":[0.5,0.5,0.5],"image_std":[0.5,0.5,0.5],"input_data_format": null,"max_frames":768,"merge_size":2,"min_frames":4,"num_frames": null,"pad_size": null,"patch_size":16,"processor_class":"Qwen3VLProcessor","resample":3,"rescale_factor":0.00392156862745098,"return_metadata": false,"size":{"longest_edge":25165824,"shortest_edge":4096},"temporal_patch_size":2,"video_metadata": null,"video_processor_type":"Qwen3VLVideoProcessor"}
[INFO|processing_utils.py:1116] 2026-02-06 17:47:50,292>> loading configuration file processor_config.json from cache at None
[INFO|processing_utils.py:1199] 2026-02-06 17:47:50,543>> Processor Qwen3VLProcessor:- image_processor: Qwen2VLImageProcessorFast {"crop_size": null,"data_format":"channels_first","default_to_square": true,"device": null,"disable_grouping": null,"do_center_crop": null,"do_convert_rgb": true,"do_normalize": true,"do_pad": null,"do_rescale": true,"do_resize": true,"image_mean":[0.5,0.5,0.5],"image_processor_type":"Qwen2VLImageProcessorFast","image_std":[0.5,0.5,0.5],"input_data_format": null,"max_pixels": null,"merge_size":2,"min_pixels": null,"pad_size": null,"patch_size":16,"processor_class":"Qwen3VLProcessor","resample":3,"rescale_factor":0.00392156862745098,"return_tensors": null,"size":{"longest_edge":16777216,"shortest_edge":65536},"temporal_patch_size":2}- tokenizer: Qwen2TokenizerFast(name_or_path='Qwen/Qwen3-VL-2B-Instruct', vocab_size=151643, model_max_length=262144, is_fast=True, padding_side='right', truncation_side='right', special_tokens={'eos_token':'<|im_end|>','pad_token':'<|endoftext|>','additional_special_tokens':['<|im_start|>','<|im_end|>','🎵','📷','🖼️','📄','📝','💬','🔤','📊']}
[INFO|trainer.py:2519] 2026-02-06 17:47:58,649>> ***** Running training *****
[INFO|trainer.py:2520] 2026-02-06 17:47:58,649>> Num examples = 600
[INFO|trainer.py:2521] 2026-02-06 17:47:58,649>> Num Epochs = 2
[INFO|trainer.py:2522] 2026-02-06 17:47:58,649>> Instantaneous batch size per device = 2
[INFO|trainer.py:2525] 2026-02-06 17:47:58,649>> Total train batch size (w. parallel, distributed & accumulation)=8
[INFO|trainer.py:2526] 2026-02-06 17:47:58,649>> Gradient Accumulation steps = 4
[INFO|trainer.py:2527] 2026-02-06 17:47:58,649>> Total optimization steps = 150
[INFO|trainer.py:2528] 2026-02-06 17:47:58,651>> Number of trainable parameters = 8,716,288
{'loss':4.3662,'grad_norm':5.828382968902588,'learning_rate':6e-06,'epoch':0.13}
{'loss':4.389,'grad_norm':6.548262119293213,'learning_rate':9.978353953249023e-06,'epoch':0.27}
{'loss':4.0005,'grad_norm':6.604191303253174,'learning_rate':9.736983212571646e-06,'epoch':0.4}
{'loss':3.4562,'grad_norm':5.726210117340088,'learning_rate':9.24024048078213e-06,'epoch':0.53}
{'loss':3.1868,'grad_norm':3.4086873531341553,'learning_rate':8.51490528712831e-06,'epoch':0.67}
{'loss':2.9764,'grad_norm':2.155060529646077e-06,'epoch':0.8}
{'loss':2.9609,'grad_norm':2.26679611206547,'learning_rate':6.545084971874738e-06,'epoch':0.93}
{'loss':2.7471,'grad_norm':1.8668205738067627,'learning_rate':5.406793373339292e-06,'epoch':1.07}
{'loss':2.9607,'grad_norm':2.023541438752585e-06,'epoch':1.2}
{'loss':2.7321,'grad_norm':1.6290875673294067,'learning_rate':3.12696703292044e-06,'epoch':1.33}
{'loss':2.6867,'grad_norm':2.182967628112793,'learning_rate':2.1083383191600676e-06,'epoch':1.47}
{'loss':2.7761,'grad_norm':1.878283852992554,'learning_rate':1.2455998350925042e-06,'epoch':1.6}
{'loss':2.6362,'grad_norm':1.8889576196670532,'learning_rate':5.852620357053651e-07,'epoch':1.73}
{'loss':2.6991,'grad_norm':2.0048000812530518,'learning_rate':1.6292390268568103e-07,'epoch':1.87}
{'loss':2.6784,'grad_norm':1.9924118518829346,'learning_rate':1.3537941026914302e-09,'epoch':2.0}
100%|███████████████████████████████████████████████████████████████████████████|150/150 [01:30<00:00,1.65it/s]
[INFO|trainer.py:4309] 2026-02-06 17:49:31,374>> Saving model checkpoint to saves/qwen3-2b-coco-3000/lora/sft/checkpoint-150
{'train_runtime':93.701,'train_samples_per_second':12.807,'train_steps_per_second':1.601,'train_loss':3.1501645787556964,'epoch':2.0}
100%|███████████████████████████████████████████████████████████████████████████|150/150 [01:31<00:00,1.63it/s]
epoch = 2.0 total_flos = 1679346GF train_loss = 3.1502 train_runtime = 0:01:33.70 train_samples_per_second = 12.807 train_steps_per_second = 1.601
Figure saved at: saves/qwen3-2b-coco-3000/lora/sft/training_loss.png
[WARNING|2026-02-06 17:49:33] llamafactory.extras.ploting:148>> No metric eval_loss to plot.
[WARNING|2026-02-06 17:49:33] llamafactory.extras.ploting:148>> No metric eval_accuracy to plot.
[INFO|modelcard.py:456] 2026-02-06 17:49:33,450>> Dropping the following result as it does not have all the necessary fields:{'task':{'name':'Causal Language Modeling','type':'text-generation'}}

训练过程记录