基于llama_factory的qwen3全参微调

最新推荐文章于 2026-01-07 20:05:05 发布

原创最新推荐文章于 2026-01-07 20:05:05 发布 · 159 阅读

0 ·

CC 4.0 BY-SA版权

文章标签：

#llama

部署运行你感兴趣的模型镜像

### model
model_name_or_path: /models/Qwen3-4B-Instruct-2507
#tool_format: qwen3
### method
stage: dpo
do_train: true
finetuning_type: full
#lora_target: all
#lora_rank: 16
pref_beta: 0.1
pref_loss: orpo  # [sigmoid (dpo), orpo, simpo]
#pissa_init: true
#pissa_iter: 16
#pissa_convert: true
deepspeed: /LLaMA-Factory-main/examples/deepspeed/ds_z3_config.json
flash_attn: fa2
### dataset
enable_thinking: false
dataset_dir: train_datas/task_dataset_info
dataset: task_file_name
template: qwen3
cutoff_len: 8000
overwrite_cache: true
preprocessing_num_workers: 16

### output
output_dir: /train_result/outputmodelname
logging_steps: 15
save_steps: 200
plot_loss: true
overwrite_output_dir: true
#report_to: swanlab
### train
per_device_train_batch_size: 2
gradient_accumulation_steps: 8
learning_rate: 5e-6
num_train_epochs: 10
lr_scheduler_type: cosine
warmup_ratio: 0.1
bf16: true
ddp_timeout: 180000000

您可能感兴趣的与本文相关的镜像