`The following are the loss images, validation set loss, and training parameters:
enter image description here
enter image description here
parameters:
model_path=/opt/workspace-cyx/model_test/Qwen1.5-14B train_dataset_dir=alpaca_gpt4_data_en,alpaca_gpt4_data_zh,oaast_sft_zh,oaast_sft
per_device_train_batch_size=4
gradient_accumulation_steps=2
output_dir=/opt/workspace-cyx/model_test/output_dir
accelerate launch --config_file accelerate_config.yaml src/train_bash.py
--max_samples 1000000
--stage sft
--do_train
--model_name_or_path ${model_path}
--dataset ${train_dataset_dir}
--template qwen
--finetuning_type lora
--lora_target q_proj,v_proj
--output_dir ${output_dir}
--per_device_train_batch_size ${per_device_train_batch_size}
--gradient_accumulation_steps ${gradient_accumulation_steps}
--lr_scheduler_type cosine
--logging_steps 5
--save_steps 2000
--learning_rate 1e-5
--num_train_epochs 1.0
--plot_loss
--fp16
--do_eval
--save_steps 100
--eval_steps 100
--val_size 0.01
--evaluation_strategy steps
Qwen and Qwen1.5 both have fine-tuning for 7B and 14B, using the four datasets alpaca_gpt4_data_en, alpaca_gpt4_data_zh, oaast_sft_zh, and oaast_sft that come with llama_factory.
Training process:
- What I initially thought was whether the parameter model was not applicable, but after trying several thousand question models, there were varying degrees of oscillations
- Then we started modifying the parameters, but when we modified batch size, lora_rank, and other parameters, the results were still almost the same
- The dataset is officially provided and there should be no problem, with a total of tens of thousands of instructions
The current idea is:
- Does this model have convergence or not? Is there no problem with model training? It’s just that Qwen1.5 has a strong ability to receive these datasets and oscillates normally [because there are no problems with the validation set]
- There is an issue with the parameters/dataset, but it has been adjusted many times and still cannot be resolved
I don’t know if anyone has encountered such problems while fine-tuning, and how they have been resolved. I hope there are someone to help me clarify,thank you!`
New contributor