It seems that compared to qwen2, there are no major changes in the model structure of mimo. I think we can resuse qwen template to do post-training or infer.

### model
model_name_or_path: XiaomiMiMo/MiMo-7B-SFT
trust_remote_code: true

### method
stage: sft
do_train: true
finetuning_type: lora
lora_rank: 8
lora_target: all

### dataset
dataset: alpaca_en_demo
template: qwen
cutoff_len: 2048
max_samples: 1000
overwrite_cache: true
preprocessing_num_workers: 16
dataloader_num_workers: 4

### output
output_dir: saves/mimo-7b/lora/sft
logging_steps: 10
save_steps: 500
plot_loss: true
overwrite_output_dir: true
save_only_model: false
report_to: none  # choices: [none, wandb, tensorboard, swanlab, mlflow]

### train
per_device_train_batch_size: 2
gradient_accumulation_steps: 8
learning_rate: 1.0e-4
num_train_epochs: 3.0
lr_scheduler_type: cosine
warmup_ratio: 0.1
fp16: true
ddp_timeout: 180000000
resume_from_checkpoint: null

BTW, please comment this line https://huggingface.co/XiaomiMiMo/MiMo-7B-SFT/blob/main/modeling_mimo.py#L64 for ddp training because it is not used during training.

Another difference is that the default system prompt is You are a helpful assistant. instead of You are Qwen, created by Alibaba Cloud. You are a helpful assistant. We can replace this in our dataset system filed.

We will update our README later.

Support Mimo #7939

Description

Reminder

Description

Pull Request

Metadata

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Issue actions