Download code/rerank_ins/scripts/train/train_qwen35.sh from WaitXian/liqi: direct link, hf CLI and curl.
- Browser
- Download file 2.39 kB
-
https://huggingface.co/WaitXian/liqi/resolve/main/code/rerank_ins/scripts/train/train_qwen35.sh
- Command line
-
hf download hf://WaitXian/liqi/code/rerank_ins/scripts/train/train_qwen35.sh
-
curl -L -o train_qwen35.sh https://huggingface.co/WaitXian/liqi/resolve/main/code/rerank_ins/scripts/train/train_qwen35.sh
2.39 kB
| # ============ Configuration ============ | |
| MODEL_ID=qwen3_5-9b | |
| MODEL_LOCAL_PATH=/mnt/workspace/siku/Qwen3.5-9B # TODO: set your model path | |
| TRAIN_DATA_PATH=/data/oss_bucket_0/siku/Last_reranker/dataset/all_data_wo_q_bbox/00-07/with_bbox/train_format/train.jsonl # TODO: set your train data path | |
| EVAL_DATA_PATH=/data/oss_bucket_0/siku/Last_reranker/dataset/all_data_wo_q_bbox/00-07/with_bbox/train_format/eval.jsonl # TODO: set your eval data path | |
| TRAIN_VISION_ENCODER=False | |
| USE_VISION_LORA=False | |
| TRAIN_VISION_PROJECTOR=False | |
| USE_LORA=False | |
| Q_LORA=False | |
| LORA_R=128 | |
| LORA_ALPHA=128 | |
| RUN_ID=${MODEL_ID}-reranker | |
| DS_STAGE=zero2 | |
| PER_DEVICE_BATCH_SIZE=2 | |
| GRAD_ACCUM=2 | |
| NUM_EPOCHS=5 | |
| LR=1e-5 | |
| MODEL_MAX_LEN=8192 | |
| NUM_TRAIN_DOCS=10 | |
| FULL_CROP_MODE=True | |
| FORMAT_PROMPT=format_prompt/listwise_reranker_v2.jinja | |
| OUTPUT_DIR=/data/oss_bucket_0/rerank_9B/output/$RUN_ID # TODO : 修改路径 | |
| # ============ Launch ============ | |
| CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 torchrun --nproc_per_node=8 train/train_reranker_qwen35.py \ | |
| --model_id $MODEL_ID \ | |
| --output_dir $OUTPUT_DIR \ | |
| --report_to tensorboard \ | |
| --run_name $RUN_ID \ | |
| --deepspeed ./ds_configs/${DS_STAGE}.json \ | |
| --bf16 True \ | |
| --num_train_epochs $NUM_EPOCHS \ | |
| --per_device_train_batch_size $PER_DEVICE_BATCH_SIZE \ | |
| --per_device_eval_batch_size 4 \ | |
| --gradient_accumulation_steps $GRAD_ACCUM \ | |
| --eval_strategy "no" \ | |
| --eval_steps 100 \ | |
| --eval_accumulation_steps 4 \ | |
| --save_strategy "steps" \ | |
| --save_steps 100 \ | |
| --save_total_limit 500 \ | |
| --learning_rate ${LR} \ | |
| --weight_decay 0. \ | |
| --warmup_ratio 0.03 \ | |
| --lr_scheduler_type "cosine" \ | |
| --logging_steps 1 \ | |
| --tf32 True \ | |
| --model_max_length $MODEL_MAX_LEN \ | |
| --gradient_checkpointing True \ | |
| --dataloader_num_workers 8 \ | |
| --train_vision_encoder $TRAIN_VISION_ENCODER \ | |
| --use_vision_lora $USE_VISION_LORA \ | |
| --train_vision_projector $TRAIN_VISION_PROJECTOR \ | |
| --use_lora $USE_LORA \ | |
| --q_lora $Q_LORA \ | |
| --lora_r $LORA_R \ | |
| --lora_alpha $LORA_ALPHA \ | |
| --model_local_path $MODEL_LOCAL_PATH \ | |
| --train_data_path $TRAIN_DATA_PATH \ | |
| --eval_data_path $EVAL_DATA_PATH \ | |
| --use_flash_attn False \ | |
| --format_prompt $FORMAT_PROMPT \ | |
| --num_train_docs $NUM_TRAIN_DOCS \ | |
| --full_crop_mode $FULL_CROP_MODE \ | |
| --max_full_tokens 160 \ | |
| --max_crop_tokens 160 | |