qwen_full_sft.sh

#!/bin/bash

CUDA_VISIBLE_DEVICES=0 python ./src/train_bash.py \
    --stage sft \
    --do_train \
    --model_name_or_path ../model/Qwen1.5-4B \
    --dataset universal_ner_all \
    --dataset_dir ./data \
    --template qwen \
    --finetuning_type full \
    --output_dir ../saves/qwen_sft/4b_full \
    --overwrite_cache \
    --overwrite_output_dir \
    --cutoff_len 3072 \
    --preprocessing_num_workers 32 \
    --per_device_train_batch_size 1 \
    --per_device_eval_batch_size 1 \
    --gradient_accumulation_steps 32 \
    --lr_scheduler_type cosine \
    --logging_steps 10 \
    --warmup_steps 200 \
    --save_steps 1000 \
    --eval_steps 1000 \
    --save_strategy steps \
    --evaluation_strategy steps \
    --load_best_model_at_end \
    --learning_rate 2e-5 \
    --num_train_epochs 1.0 \
    --val_size 0.1 \
    --plot_loss \
    --report_to wandb