二次预训练阶段全参微调,损失曲线是否正常，如何优化

[Shame-fight]

Reminder

• I have read the README and searched the existing issues.

System Info

llamafactory== 0.9.1.dev0

Reproduction

model

model_name_or_path: /nanshu_data/jgx/LLM_Model/Qwen/Qwen2___5-0___5B

method

stage: pt
do_train: true
finetuning_type: full
deepspeed: examples/deepspeed/ds_z3_config.json

dataset

dataset: yingji_pt,CCI3_2w,refineweb,wiki_en,wiki_ch_6k,alpha_gpt4,BELLE
template: default
cutoff_len: 1024
max_samples: 200000
overwrite_cache: true
preprocessing_num_workers: 16

output

output_dir: saves/qwen2.5-0.5b/full/pt
logging_steps: 10
save_steps: 1000
plot_loss: true
overwrite_output_dir: true

train

per_device_train_batch_size: 2
gradient_accumulation_steps: 4
learning_rate: 1.0e-5
num_train_epochs: 4.0
lr_scheduler_type: cosine
warmup_ratio: 0.1
bf16: true
ddp_timeout: 180000000

eval

val_size: 0.05
per_device_eval_batch_size: 1
eval_strategy: steps
eval_steps: 1000

Expected behavior

预训练阶段train loss阶梯式下降，eval loss先下降后上升，是否正常。5000个step后出现过拟合？请问是否需要继续预训练，如何改进。指令监督微调应该选择多少step的二次预训练模型

Others

No response

[aHumbleblackcat]

您好，我也遇到了类似问题，方便加个联系方式交流一下吗！vx：18923400893