环境

单卡 A100

DataSet

数据清洗, 详细看文档, 关注下数据的格式

Distill_data_17k-train.arrow

DataSet 配置

在LLama-Factory中注册自定义数据集，找到dataset_info.json

"Distill": {
    "file_name": "/data/jupyterfile/wei/distill/Distill_data_17k-train.arrow",
    "formatting": "sharegpt",
    "columns": {
        "messages": "conversations",
        "system": "system"
    },
    "tags": {
        "role_tag": "from",
        "content_tag": "value",
        "user_tag": "user",
        "assistant_tag": "assistant"
    }
}

Train 脚本

【配置 qwen2_full_sft.yaml】

### model
model_name_or_path: /data/jupyterfile/wei/distill/Qwen2.5-1.5B-Instruct

### method
stage: sft
do_train: true
finetuning_type: full
deepspeed: examples/deepspeed/ds_z3_offload_config.json  # choices: [ds_z0_config.json, ds_z2_config.json, ds_z3_config.json]

### dataset
dataset: Distill
template: qwen
cutoff_len: 8192
max_samples: 100000
overwrite_cache: true
preprocessing_num_workers: 16

### output
output_dir: ./saves/MiniDeepSeekR1/full/original
logging_steps: 1
save_steps: 100
plot_loss: true

### train
per_device_train_batch_size: 1
gradient_accumulation_steps: 12
learning_rate: 1.0e-5
num_train_epochs: 1.0 #3.0 
lr_scheduler_type: cosine
warmup_ratio: 0.1
bf16: true
ddp_timeout: 180000000

report_to: wandb
run_name: Distill

Train