bf16: true
cutoff_len: 6000
dataset: critic_training_data,refined,omnisvg,svgen,llm4svg
dataset_dir: data
ddp_timeout: 180000000
deepspeed: cache/ds_z3_config.json
do_train: true
eval_steps: 5000
eval_strategy: steps
finetuning_type: full
flash_attn: auto
gradient_accumulation_steps: 8
include_num_input_tokens_seen: true
learning_rate: 0.0001
logging_steps: 21
lr_scheduler_type: cosine
max_grad_norm: 1.0
max_samples: 1000000
model_name_or_path: /gemini-3/space/thu/zhaozhiyuan/wfy-mptsnet/Qwen2.5-VL-7B-Instruct
num_train_epochs: 3.0
optim: adamw_torch
output_dir: saves/Qwen2.5-VL-7B-Instruct/full/all_stage_123
packing: false
per_device_eval_batch_size: 2
per_device_train_batch_size: 2
plot_loss: true
preprocessing_num_workers: 16
report_to: none
save_steps: 5000
stage: sft
template: qwen2_vl
trust_remote_code: true
val_size: 0.03
warmup_steps: 0