Qwen-2.5-32B base 模型Lora SFT之后输出后续会产生乱码

[zzh-SJTU]

用Lora SFT Qwen-2.5-32B， 数据用的的是 Tulu-3 (https://huggingface.co/datasets/allenai/tulu-3-sft-mixture), 数据量12K:

model

model_name_or_path: Qwen/Qwen2.5-32B

method

stage: sft
do_train: true
finetuning_type: lora
lora_target: all
deepspeed: examples/deepspeed/ds_z3_config.json

dataset

dataset: tulu_12000
template: qwen
cutoff_len: 2048
overwrite_cache: true
preprocessing_num_workers: 16

output

output_dir: saves/Qwen2.5-32B_tulu_12000/lora/sft
logging_steps: 10
save_steps: 100000
plot_loss: true
overwrite_output_dir: true
report_to: wandb
run_name: run_Qwen2.5-32B_tulu_12000# optional

train

per_device_train_batch_size: 1
gradient_accumulation_steps: 2
learning_rate: 1.0e-5
num_train_epochs: 3.0
lr_scheduler_type: cosine
warmup_ratio: 0.1
bf16: true
ddp_timeout: 180000000

eval

val_size: 0.01
per_device_eval_batch_size: 1
eval_strategy: steps
eval_steps: 1000

[zzh-SJTU]

例如回答数学问题在完成回答之后会产生乱码

[hiyouga]

base 模型用 template default

[zzh-SJTU]

感谢回答！理论上来说只要training和inference用同一个template就可以，为什么要用default呢？