38 lines
1.0 KiB
Bash
38 lines
1.0 KiB
Bash
# pip install flash-attn --no-build-isolation
|
|
|
|
|
|
nproc_per_node=1
|
|
|
|
MLFLOW_TRACKING_URI=http://0.0.0.0:5004 \
|
|
INFONCE_USE_BATCH=True \
|
|
CUDA_VISIBLE_DEVICES=0 \
|
|
NPROC_PER_NODE=$nproc_per_node \
|
|
swift sft \
|
|
--model $(pwd)/../../data/models/Qwen3-Embedding-0.6B/model \
|
|
--task_type embedding \
|
|
--model_type qwen3_emb \
|
|
--train_type lora \
|
|
--lora_rank 16 \
|
|
--lora_alpha 32 \
|
|
--target_modules all-linear \
|
|
--max_length 2048 \
|
|
--dataset v11_dataset_hn \
|
|
--custom_register_path $(pwd)/../../data/dataset/v11_dataset_hn/generated.py \
|
|
--split_dataset_ratio 0.005 \
|
|
--eval_strategy steps \
|
|
--output_dir output \
|
|
--eval_steps 1000 \
|
|
--num_train_epochs 1 \
|
|
--save_steps 1000 \
|
|
--save_total_limit 10 \
|
|
--per_device_train_batch_size 16 \
|
|
--per_device_eval_batch_size 16 \
|
|
--gradient_accumulation_steps 4 \
|
|
--learning_rate 3.0e-6 \
|
|
--lr_scheduler_type constant \
|
|
--loss_type infonce \
|
|
--label_names labels \
|
|
--dataloader_drop_last true \
|
|
--deepspeed zero3 \
|
|
--report_to mlflow
|