Datasets:
MLVU
/

Languages:
English
ArXiv:
License:

caption anno file

#1
by gogorunrun - opened

Where is the answer to the test_generation_tasks.json file in this dataset?

Hello,

At the current stage, the ground truth for the MLVU test is not publicly available.

You can send your result files to zhoujunjie[#]bupt.edu.cn (replace [#] with @).

For multiple-choice questions, please organize your result files according to the guidelines at https://github.com/JUNJIE99/MLVU/tree/main/evaluation_test.

For generative tasks, please follow these rules when organizing your files:

  • Separate the summary and subscene captioning into two JSON files.
  • For each question, use the keys "video_name", "Q", and "pred" to store the video name, question, and your model's prediction, respectively.

There is no submission limit for multiple-choice questions. However, for generative tasks, due to the cost associated with GPT usage, we provide a maximum of one evaluation per user. Therefore, please ensure that your model is well-optimized on the dev set before submitting for the test set generative tasks.

P.S. For evaluating multiple-choice questions on the MLVU dev set, we recommend using lmms-eval, which allows you to complete the MLVU dev set evaluation with just one click.

Thanks for the reply.

Sign up or log in to comment