- WARNING - llmtuner.extras.ploting - No metric eval_loss to plot.

[kc-two]

After training LLama2-7b-chat, I get only train_loss values and plot.

Can we also get eval_loss values and plot through this training?

Anybody has already implemented?

[hiyouga]

See https://github.com/hiyouga/LLaMA-Factory/blob/main/examples/lora_single_gpu/sft.sh

[kc-two]

I have one more question. I followed the example provided, but I'm not seeing any information about the eval_loss on the plot. I can see the eval loss values in the log, but they don't appear on the plot. Do you have any advice on this matter? Thank you for your help.

** Train loss plot works great.

command:
CUDA_VISIBLE_DEVICES=1 python src/train_bash.py
--stage sft
--do_train
--model_name_or_path “ — Llama-2-3237b-chat-hf-ko3r_ver1"
--dataset eng_kr_train,eng_kr_eval
--dataset_dir data
--template vanilla
--finetuning_type lora
--lora_target q_proj,k_proj,v_proj,o_proj
--output_dir saves/LLaMjojA332-ko-7b/lora/3332
--overwrite_cache
--overwrite_output_dir
--cutoff_len 1024
--preprocessing_num_workers 16
--per_device_train_batch_size 4
--per_device_eval_batch_size 4
--gradient_accumulation_steps 8
--lr_scheduler_type constant
--logging_steps 10
--warmup_steps 20
--save_steps 100
--eval_steps 100
--evaluation_strategy steps
--load_best_model_at_end
--learning_rate 2e-4
--num_train_epochs 1.0
--max_samples 3000
--val_size 0.1
--plot_loss
--fp16

[hiyouga]

your training step is too small to plot a eval loss, use --eval_steps 10

[kc-two]

I greatly appreciate your advice and help!