| NPROC_PER_NODE=2 \ | |
| CUDA_VISIBLE_DEVICES=0,1 \ | |
| MAX_NEGATIVE_SAMPLES=1 \ | |
| swift sft \ | |
| --model JinaAI/jina-reranker-m0 \ | |
| --task_type reranker \ | |
| --loss_type pointwise_reranker \ | |
| --tuner_type lora \ | |
| --dataset swift/TextCaps:rerank \ | |
| --split_dataset_ratio 0.05 \ | |
| --eval_strategy steps \ | |
| --output_dir output \ | |
| --eval_steps 100 \ | |
| --num_train_epochs 1 \ | |
| --save_steps 200 \ | |
| --per_device_train_batch_size 4 \ | |
| --per_device_eval_batch_size 4 \ | |
| --gradient_accumulation_steps 8 \ | |
| --dataset_num_proc 8 \ | |
| --learning_rate 6e-5 \ | |
| --label_names labels \ | |
| --dataloader_drop_last true \ | |
| --attn_impl flash_attn \ | |
| --padding_free true | |