pixel_reasoner

Pixel Reasoner

Data Preprocessing

python examples/data_preprocess/pixel_reasoner/prepare_train.py --dataset_path=TIGER-Lab/PixelReasoner-RL-Data --local_dir=data/pixel_reasoner --version max_8192 --include_videos=True --filter_len=8192
python examples/data_preprocess/pixel_reasoner/prepare_train.py --dataset_path=TIGER-Lab/PixelReasoner-RL-Data --local_dir=data/pixel_reasoner --version max_16384 --include_videos=True --filter_len=16384
python examples/data_preprocess/pixel_reasoner/prepare_train.py --dataset_path=TIGER-Lab/PixelReasoner-RL-Data --local_dir=data/pixel_reasoner --include_videos=True

note the data processing will filter out those samples with length larger than 8192 and will take a while (0.5 to 1 hour) to finish. If you don't want to filter, remove the --filter_len argument. But there are some samples with length larger than 8192, which may cause problems in training, so make sure you set the max_prompt_length during the training properly.

Training

bash examples/train/pixel_reasoner/train_qwen25vl.sh

It should be able to run under 8 H100/A100 GPUs with 80GB memory.

Tips:

• if output shared memory, try lower the data.dataloader_num_workers
• if out of cuda memory during vllm rollout, try set actor_rollout_ref.rollout.enforce_eager=True, might be slower.
• if out of cuda memory during training, try lower the use_dynamic_bs=False.

Evaluation

1. Dataset

python examples/data_preprocess/pixel_reasoner/infovqa.py --dataset_path=JasperHaozhe/InfoVQA-EvalData-PixelReasoner --split=test --local_dir=data/pixel_reasoner/info_vqa
python examples/data_preprocess/pixel_reasoner/mvbench.py --dataset_path=JasperHaozhe/MVBench-EvalData-PixelReasoner --split=test --local_dir=data/pixel_reasoner/mvbench
python examples/data_preprocess/pixel_reasoner/tallyqa.py --dataset_path=JasperHaozhe/TallyQA-EvalData-PixelReasoner --split=test --local_dir=data/pixel_reasoner/tallyqa
python examples/data_preprocess/pixel_reasoner/vstar.py --dataset_path=JasperHaozhe/VStar-EvalData-PixelReasoner --split=test --local_dir=data/pixel_reasoner/vstar

2. Evaluation

bash examples/train/pixel_reasoner/eval.sh

Notes

• If you kill a training job, for current version of verl, it seems the gpu memory will not be released immediately. You may need to kill mannually (e.g. pkill -f -9 ray)
• Check original paper for more details: Pixel Reasoner
• PR
• requires transformers<4.53.0 due to Qwen2.5-VL model codes