微调后的模型无法输出特殊token

[THUchenzhou]

任务：使用大模型做简单的文本二分类。
使用如下命令添加special_tokens，微调qwen
CUDA_VISIBLE_DEVICES=0,1 llamafactory-cli train examples/train_lora/qwen2_special_token.yaml

model_name_or_path: /data/Qwen/Qwen2.5-0.5B-Instruct

stage: sft
do_train: true
finetuning_type: lora
lora_target: all
deepspeed: examples/deepspeed/ds_z0_config.json
lora_rank: 16
lora_alpha: 32

dataset: special_token
template: qwen
cutoff_len: 4096
overwrite_cache: true
preprocessing_num_workers: 16

output_dir: saves/lora/sft/12_general/qwen25-0.5b/tag_15761_wo_lora_16_lr_500_special_token
logging_steps: 10
save_steps: 6000
plot_loss: true
overwrite_output_dir: true

per_device_train_batch_size: 4
gradient_accumulation_steps: 1
learning_rate: 5.00e-4
num_train_epochs: 1
lr_scheduler_type: cosine
warmup_ratio: 0.1
bf16: true
ddp_timeout: 180000000
new_special_tokens: "<|LABEL_0|>,<|LABEL_1|>"
resize_vocab: true
additional_target: embed_tokens,lm_head

启动服务CUDA_VISIBLE_DEVICES=0 API_PORT=8000 llamafactory-cli api examples/inference/qwen2.yaml，文件内容如下：

model_name_or_path:/data/Qwen/Qwen2.5-0.5B-Instruct
adapter_name_or_path: saves/lora/sft/12_general/qwen25-0.5b/tag_15761_wo_lora_16_lr_500_special_token
template: qwen
finetuning_type: lora

问题：
启动服务后，请求微调后的模型，无法输出新添加的特殊token。请问是训练参数设置错误，还是启动服务的方式错误？

[hiyouga]

添加参数 skip_special_tokens: true

[BeSkyer]

添加参数 skip_special_tokens: true

我也这样添加了，但是没有作用

model

model_name_or_path: /home/huggingface_down/Qwen/Qwen2.5-7B-Instruct

#deepspeed: /home/MoE_LLM/LLaMA-Factory/examples/deepspeed/ds_z2_config.json

flash_attn: auto

method

stage: sft
do_train: true
finetuning_type: lora
lora_target: all
lora_rank: 32
lora_alpha: 64
lora_dropout: 0.03

dataset

dataset: sft_MetaMathQA-balanced-20k-cleaned
template: qwen
cutoff_len: 1024
overwrite_cache: true
preprocessing_num_workers: 8

output

output_dir: /home/MoE_LLM/saves_SFT/qwen2-7b/lora3_MetaMathQA-20k_true/
logging_steps: 10
save_steps: 500
plot_loss: true
overwrite_output_dir: false

train

per_device_train_batch_size: 1
gradient_accumulation_steps: 4
learning_rate: 1.0e-5
num_train_epochs: 5.0 #epochs
lr_scheduler_type: cosine
warmup_ratio: 0.03
bf16: true
ddp_timeout: 180000000
new_special_tokens : ",,,"
resize_vocab : true
additional_target : embed_tokens,lm_head
skip_special_tokens: true

eval

val_size: 0.02
per_device_eval_batch_size: 1

eval_strategy: steps

eval_strategy: steps
eval_steps: 50000
#resume_from_checkpoint : true

[hiyouga]

@BeSkyer Fixed in #6624, please update code