[help] Weird loss qwen3-vl-8b

[Damon-GSY]

PROMPT_TEMPLATE="qwen3_vl_nothink"
MODEL_NAME="Qwen/Qwen3-VL-8B-Instruct"

args="--stage sft \
    --model_name_or_path=$MODEL_NAME \
    --do_train \
    --template=${PROMPT_TEMPLATE} \
    --finetuning_type full \
    --output_dir=local/tmp/ckpt_save_path/ \
    --overwrite_cache \
    --per_device_train_batch_size 1 \
    --gradient_accumulation_steps 4 \
    --num_train_epochs 2 \
    --lr_scheduler_type cosine \
    --logging_steps 10 \
    --save_steps 200 \
    --learning_rate=$LR \
    --cutoff_len=2048 \
    --preprocessing_num_workers=8 \
    --dataloader_num_workers=8 \
    --plot_loss \
    --report_to=wandb \
    --deepspeed=scripts/ds_zero3_cpuoffload.json \
    --flash_attn fa2 \
    --dataset_name=seasonality_combined_alpaca\
    --prompt=instruction \
    --query=input \
    --response=output \
    --streaming \
    --max_steps=12000 \
    --image=image \
    --image_resolution 12000 \
    --gradient_checkpointing \
    --bf16"

Used 8 * H100.
Recently I tried to do SFT on qwen3-vl-8b. I prepare a dataset(200,000 data, around 160,000 task specific data and 40,000 sharegpt4v data)

then I come across this loss

• I’m not sure why, between steps 2810 and 2820, the loss suddenly drops from 0.63 to 0.45.
• I also don’t understand why the loss curve shows a distinct step-like drop pattern.

I have verified that the training data is being shuffled properly, and one epoch takes roughly 6,000 steps.

[hiyouga]

Because you are using streaming, the dataset was not fully shuffled. You can shuffle the data manually before training

[Damon-GSY]

I have double checked that the data is fully shuffled before I run the training scripts

[hiyouga]

Can you print the data example from the 2 different steps?

[Damon-GSY]

I don't see if the wandb correctly log the trace. Any there any arguments of LLaMA-Factory to set this up and also configure the validation set?

[hiyouga]

You can modify the codebase to print the data example