1!python qlora.py \
2 --model_name cyberagent/calm2-7b-chat \
3 --output_dir "./output/calm2-7b-chat-tsukuyomi-chan" \
4 --dataset "alpaca" \
5 --max_steps 1500 \
6 --use_auth \
7 --logging_steps 100 \
8 --save_strategy steps \
9 --data_seed 42 \
10 --save_steps 300 \
11 --save_total_limit 100 \
12 --max_new_tokens 5000 \
13 --dataloader_num_workers 1 \
14 --group_by_length \
15 --logging_strategy steps \
16 --remove_unused_columns False \
17 --do_train \
18 --lora_r 64 \
19 --lora_alpha 16 \
20 --lora_modules all \
21 --double_quant \
22 --quant_type nf4 \
23 --bf16 \
24 --bits 4 \
25 --warmup_ratio 0.03 \
26 --lr_scheduler_type constant \
27 --gradient_checkpointing \
28 --source_max_len 1024 \
29 --target_max_len 512 \
30 --per_device_train_batch_size 4 \
31 --gradient_accumulation_steps 4 \
32 --eval_steps 187 \
33 --learning_rate 0.0001 \
34 --adam_beta2 0.999 \
35 --max_grad_norm 0.3 \
36 --lora_dropout 0.1 \
37 --weight_decay 0.0 \
38 --seed 0 \
39 --load_in_4bit \
40 --use_peft \
41 --batch_size 4 \
42 --gradient_accumulation_steps 2
