chore: import upstream snapshot with attribution

examples/train/reranker/qwen3/infer.py

@@ -0,0 +1,40 @@
+# For full-parameter training, please refer to:
+# https://github.com/modelscope/ms-swift/blob/main/examples/infer/demo_reranker.py
+
+from swift.infer_engine import InferRequest, TransformersEngine
+
+
+def run_qwen3_reranker():
+    engine = TransformersEngine(
+        'Qwen/Qwen3-Reranker-4B',
+        task_type='generative_reranker',
+        attn_impl='flash_attention_2',
+        adapters=['output/vx-xxx/checkpoint-xxx'])
+
+    infer_requests = [
+        InferRequest(messages=[{
+            'role': 'user',
+            'content': 'Mindful emotion regulation: An integrative review.'
+        }, {
+            'role':
+            'assistant',
+            'content':
+            'Differential effects of mindful breathing, progressive muscle relaxation, and loving-kindness '
+            'meditation on decentering and negative reactions to repetitive thoughts.'
+        }]),
+        InferRequest(messages=[{
+            'role': 'user',
+            'content': 'Mindful emotion regulation: An integrative review.'
+        }, {
+            'role': 'assistant',
+            'content': 'Exploiting vulnerability to secure user privacy on a social networking site'
+        }])
+    ]
+
+    responses = engine.infer(infer_requests)
+    scores = [response.choices[0].message.content for response in responses]
+    print(f'scores: {scores}')
+
+
+if __name__ == '__main__':
+    run_qwen3_reranker()

examples/train/reranker/qwen3/qwen3_reranker.sh

@@ -0,0 +1,35 @@
+# 2*20GiB
+# losses: swift/loss
+CUDA_VISIBLE_DEVICES=0,1 \
+NPROC_PER_NODE=2 \
+swift sft \
+    --model Qwen/Qwen3-Reranker-4B \
+    --task_type generative_reranker \
+    --loss_type pointwise_reranker \
+    --tuner_type lora \
+    --lora_rank 8 \
+    --lora_alpha 32 \
+    --learning_rate 5e-5 \
+    --target_modules all-linear \
+    --dataset MTEB/scidocs-reranking \
+    --attn_impl flash_attn \
+    --padding_free true \
+    --torch_dtype bfloat16 \
+    --load_from_cache_file true \
+    --split_dataset_ratio 0.02 \
+    --eval_strategy steps \
+    --output_dir output \
+    --save_steps 50 \
+    --eval_steps 50 \
+    --save_total_limit 2 \
+    --logging_steps 5 \
+    --num_train_epochs 1 \
+    --max_length 4096 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --gradient_accumulation_steps 8 \
+    --dataloader_num_workers 4 \
+    --dataset_num_proc 4 \
+    --warmup_ratio 0.05 \
+    --dataloader_drop_last true \
+    --deepspeed zero2

examples/train/reranker/qwen3/qwen3_vl_reranker.sh

@@ -0,0 +1,35 @@
+# 2*70GiB
+# losses: swift/loss
+CUDA_VISIBLE_DEVICES=0,1 \
+NPROC_PER_NODE=2 \
+swift sft \
+    --model Qwen/Qwen3-VL-Reranker-8B \
+    --task_type generative_reranker \
+    --loss_type pointwise_reranker \
+    --tuner_type lora \
+    --lora_rank 8 \
+    --lora_alpha 32 \
+    --learning_rate 5e-5 \
+    --target_modules all-linear \
+    --dataset swift/TextCaps:rerank \
+    --attn_impl flash_attn \
+    --padding_free true \
+    --torch_dtype bfloat16 \
+    --load_from_cache_file true \
+    --split_dataset_ratio 0.02 \
+    --eval_strategy steps \
+    --output_dir output \
+    --save_steps 50 \
+    --eval_steps 50 \
+    --save_total_limit 2 \
+    --logging_steps 5 \
+    --num_train_epochs 1 \
+    --max_length 4096 \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --gradient_accumulation_steps 8 \
+    --dataloader_num_workers 4 \
+    --dataset_num_proc 4 \
+    --warmup_ratio 0.05 \
+    --dataloader_drop_last true \
+    --deepspeed zero2

examples/train/reranker/train_generative_reranker.sh

@@ -0,0 +1,25 @@
+# 4*47G
+# losses: swift/loss
+CUDA_VISIBLE_DEVICES=0,1,2,3 \
+NPROC_PER_NODE=4 \
+swift sft \
+    --model Qwen/Qwen3-Reranker-4B \
+    --task_type generative_reranker \
+    --loss_type pointwise_reranker \
+    --tuner_type full \
+    --dataset MTEB/scidocs-reranking \
+    --load_from_cache_file true \
+    --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 2 \
+    --per_device_eval_batch_size 2 \
+    --gradient_accumulation_steps 8 \
+    --dataset_num_proc 8 \
+    --learning_rate 6e-6 \
+    --label_names labels \
+    --deepspeed zero2 \
+    --dataloader_drop_last true

examples/train/reranker/train_generative_reranker_listwise.sh

@@ -0,0 +1,25 @@
+# 4*47G
+# losses: swift/loss
+CUDA_VISIBLE_DEVICES=0,1,2,3 \
+NPROC_PER_NODE=4 \
+swift sft \
+    --model Qwen/Qwen3-Reranker-4B \
+    --task_type generative_reranker \
+    --loss_type listwise_reranker \
+    --tuner_type full \
+    --dataset MTEB/scidocs-reranking \
+    --load_from_cache_file true \
+    --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 2 \
+    --per_device_eval_batch_size 2 \
+    --gradient_accumulation_steps 8 \
+    --dataset_num_proc 8 \
+    --learning_rate 6e-6 \
+    --label_names labels \
+    --deepspeed zero2 \
+    --dataloader_drop_last true

examples/train/reranker/train_reranker.sh

@@ -0,0 +1,22 @@
+# 1*5G
+CUDA_VISIBLE_DEVICES=0 \
+swift sft \
+    --model iic/gte-reranker-modernbert-base \
+    --task_type reranker \
+    --loss_type pointwise_reranker \
+    --tuner_type full \
+    --dataset MTEB/scidocs-reranking \
+    --load_from_cache_file true \
+    --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 64 \
+    --per_device_eval_batch_size 64 \
+    --gradient_accumulation_steps 1 \
+    --dataset_num_proc 8 \
+    --learning_rate 6e-6 \
+    --label_names labels \
+    --dataloader_drop_last true

examples/train/reranker/train_reranker_auto_patch.sh

@@ -0,0 +1,25 @@
+CUDA_VISIBLE_DEVICES=0 \
+MAX_NEGATIVE_SAMPLES=1 \
+MAX_PIXELS=602112 \
+swift sft \
+    --model Qwen/Qwen2.5-VL-3B-Instruct \
+    --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 8 \
+    --per_device_eval_batch_size 8 \
+    --gradient_accumulation_steps 4 \
+    --padding_free true \
+    --attn_impl flash_attn \
+    --dataset_num_proc 8 \
+    --learning_rate 6e-5 \
+    --label_names labels \
+    --dataloader_drop_last true \
+    --attn_impl flash_attn

examples/train/reranker/train_reranker_listwise.sh

@@ -0,0 +1,22 @@
+# 1*5G
+CUDA_VISIBLE_DEVICES=0 \
+swift sft \
+    --model iic/gte-reranker-modernbert-base \
+    --task_type reranker \
+    --loss_type listwise_reranker \
+    --tuner_type full \
+    --dataset MTEB/scidocs-reranking \
+    --load_from_cache_file true \
+    --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 64 \
+    --per_device_eval_batch_size 64 \
+    --gradient_accumulation_steps 1 \
+    --dataset_num_proc 8 \
+    --learning_rate 6e-6 \
+    --label_names labels \
+    --dataloader_drop_last true \

examples/train/reranker/train_reranker_mm.sh

@@ -0,0 +1,24 @@
+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