bf16: true
cutoff_len: 3000
dataset: pumlGenV3
dataset_dir: data
ddp_timeout: 180000000
deepspeed: llamaboard_cache/ds_z3_config.json
do_train: true
enable_thinking: false
finetuning_type: full
flash_attn: auto
freeze_multi_modal_projector: true
freeze_vision_tower: true
gradient_accumulation_steps: 16
image_max_pixels: 589824
image_min_pixels: 1024
include_num_input_tokens_seen: true
learning_rate: 5.0e-05
logging_steps: 1
lr_scheduler_type: cosine
max_grad_norm: 1.0
max_samples: 100000
model_name_or_path: Qwen/Qwen3.5-4B-Base
num_train_epochs: 3.0
optim: paged_adamw_8bit
output_dir: saves/Qwen3.5-4B-Base/full/Qwen3.5-4B-Base-PumlGenV3
packing: false
per_device_train_batch_size: 1
plot_loss: true
preprocessing_num_workers: 16
report_to: wandb
save_steps: 2500
stage: sft
template: qwen3_5_nothink
trust_remote_code: true
video_max_pixels: 65536
video_min_pixels: 256
warmup_steps: 0