constants.py

# this is .py for store constants 
MODEL_INFO = ["Model Type", "Model", "Language Model", "Evaluation Method"]
MODEL_INFO_V2 = ["Model", "Language Model", "Evaluation Method"]
MODEL_SIZE = ["<10B", ">=10B", "-"]
EVALUATION_METHOD = ["PPL", "PPL for A/B/C/D", "Generate", "NG"]
DIMENSION_LEVEL = ["L1", "L2", "L3"]
LEADERBOARD_VERSION = ["Version1", "Version2"]
TASK_INFO = ["Avg. All", "Avg. Img", "Avg. Video", "Scene Understanding", "Instance Identity", "Instance Attribute", "Instance Location", "Instance Counting", "Spatial Relation", "Instance Interaction", "Visual Reasoning", "Text Recognition",  "Action Recognition", "Action Prediction", "Procedure Understanding"]
TASK_V2_INFO = ["Avg. P1", "Avg. P2", "Avg. P3", "Scene Understanding", "Instance Identity", "Instance Attribute", "Instance Location", "Instance Counting", "Spatial Relation", "Instance Interaction", "Visual Reasoning", "Text Recognition", "Celebrity Recognition", "Landmark Recognition", "Chart Understanding", "Visual Referring Expression", "Science Knowledge", "Emotion Recognition", "Visual Mathematics", "Difference Spotting", "Meme Comprehension", "Global Video Understanding", "Action Recognition", "Action Predicion", "Procedure Understanding", "In-Context Captioning", "Interleaved Image-Text Analysis", "Text-to-Image Generation", "Next Image Prediction", "Text-Image Creation"]

AVG_INFO = ["Avg. All", "Avg. Img", "Avg. Video"]
AVG_V2_INFO = ["Avg. P1", "Avg. P2", "Avg. P3"]

DATA_TITILE_TYPE = ["markdown", "markdown", "markdown", "markdown", "number", "number", "number", "number", "number", "number", "number", "number", "number", "number", "number", "number", "number", "number", "number"]
DATA_TITILE_V2_TYPE = ["markdown", "markdown","markdown", "number", "number", "number", "number", "number", "number", "number", "number", "number", "number", "number", "number", "number", "number", "number", "number", "number", "number", "number", "number", "number", "number", "number", "number", "number", "number", "number", "number", "number", "number"]
CSV_DIR = "./file/result.csv"
CSV_V2_DIR = "./file/result_v2.csv"

COLUMN_NAMES = MODEL_INFO + TASK_INFO
COLUMN_V2_NAMES = MODEL_INFO_V2 + TASK_V2_INFO

DATA_NUM = [3158, 1831, 4649, 978, 2447, 657, 97, 331, 85, 1740, 2077, 1192]
DATA_NUM_V2 = [3158, 1831, 4649, 978, 2447, 657, 97, 331, 435, 330, 500, 501, 199, 277, 501, 132, 501, 159, 1594, 1509, 1225, 1023, 120, 49, 1008, 81, 79]

LEADERBORAD_INTRODUCTION = """# SEED-Bench Leaderboard
    
    Welcome to the leaderboard of the SEED-Bench! 🏆 

    SEED-Bench-1 consists of 19K multiple-choice questions with accurate human annotations for evaluating Multimodal LLMs, covering 12 evaluation dimensions including both the spatial and temporal understanding.
    Please refer to [SEED-Bench-1 paper](https://arxiv.org/abs/2307.16125) for more details.

    SEED-Bench-2 comprises 24K multiple-choice questions with accurate human anno- tations, which spans 27 dimensions, including the evalu- ation of both text and image generation.
    Please refer to [SEED-Bench-2 paper](https://arxiv.org/abs/2311.17092) for more details.
    """

SUBMIT_INTRODUCTION = """# Submit on SEED Benchmark Introduction
    1. Obtain JSON file from our [github repository](https://github.com/AILab-CVC/SEED-Bench#leaderboard-submit) after evaluation. For example on SEED-Bench-1, you can obtain InstructBLIP's JSON file as results/results.json after running 
    ```shell
    python eval.py --model instruct_blip --anno_path SEED-Bench.json --output-dir results
    ```
    And for example on SEED-Bench-2, you can obtain InternLM_Xcomposer_VL's JSON file as results/results.json after running 
    ```shell
    python eval.py --model InternLM_Xcomposer_VL --anno_path SEED-Bench_v2_level1_2_3.json --output-dir results --evaluate_level L2 --evaluate_part all --evaluate_version v2
    ```
    2. If you want to update model performance by uploading new results, please ensure 'Model Name Revision' is the same as what's shown in the leaderboard. For example, if you want to modify InstructBLIP's performance, you need to fill in 'InstructBLIP' in 'Revision Model Name'.
    3. Please provide the correct link of your model's repository for each submission.  
    4. For the evaluation dimension, you can choose "All/Image/Video" for SEED-Bench-1 and "L1/L2/L3" for SEED-Bench-2, and the results of dimensions that are not evaluated will be set to zero.
    5. After clicking 'Submit Eval', you can click 'Refresh' to obtain the latest result in the leaderboard.

    ## Submit Example
    For example on SEED-Bench-1, if you want to upload InstructBLIP's result in the leaderboard, you need to:
    1. Fill in 'InstructBLIP' in 'Model Name' if it is your first time to submit your result (You can leave 'Revision Model Name' blank).
    2. Fill in 'InstructBLIP' in 'Revision Model Name' if you want to update your result (You can leave 'Model Name' blank).
    3. Select 'ImageLLM' in 'Model Type'.
    4. Fill in 'https://github.com/salesforce/LAVIS' in 'Model Link'.
    5. Fill in '7B' in 'Model size'.
    6. Select 'v1' in 'Benchmark version'.
    7. Select 'Flan-T5-XL' in 'LLM Type'.
    8. Select 'All' in 'Evaluation Dimension for SEED-Bench 1'.
    9. Select 'PPL' in 'Evaluate Method'.
    10. Upload results.json.
    11. Click the 'Submit Eval' button.
    12. Click 'Refresh' to obtain the uploaded leaderboard.

    For example on SEED-Bench-2, if you want to upload InternLM_Xcomposer_VL's result in the leaderboard, you need to:
    1. Fill in 'InternLM_Xcomposer_VL' in 'Model Name' if it is your first time to submit your result (You can leave 'Revision Model Name' blank).
    2. Fill in 'InternLM_Xcomposer_VL' in 'Revision Model Name' if you want to update your result (You can leave 'Model Name' blank).
    3. Select 'ImageLLM' in 'Model Type'.
    4. Fill in 'https://github.com/InternLM/InternLM-XComposer' in 'Model Link'.
    5. Fill in '7B' in 'Model size'.
    6. Select 'v2' in 'Benchmark version'.
    7. Select 'Other' in 'LLM Type'.
    8. Fill 'InternLM-7B' in 'LLM model(for Other)'
    9. Select 'L2' in 'Evaluation Dimension for SEED-Bench 2'.
    10. Select 'PPL' in 'Evaluate Method'.
    11. Upload results.json.
    12. Click the 'Submit Eval' button.
    13. Click 'Refresh' to obtain the uploaded leaderboard.
"""

TABLE_INTRODUCTION = """In the table below, we summarize each task performance of all the models.
        We use accurancy(%) as the primary evaluation metric for each tasks. 

        SEED-Bench-1 calculates the overall accuracy by dividing the total number of correct QA answers by the total number of QA questions.

        SEED-Bench-2 represents the overall accuracy using the average accuracy of each dimension.

        For PPL evaluation method, we count the loss for each candidate and select the lowest loss candidate. For detail, please refer [InternLM_Xcomposer_VL_interface](https://github.com/AILab-CVC/SEED-Bench/blob/387a067b6ba99ae5e8231f39ae2d2e453765765c/SEED-Bench-2/model/InternLM_Xcomposer_VL_interface.py#L74).

        For PPL A/B/C/D evaluation method, please refer [EVAL_SEED.md](https://github.com/QwenLM/Qwen-VL/blob/master/eval_mm/seed_bench/EVAL_SEED.md) for more information.

        For Generate evaluation method, please refer [Evaluation.md](https://github.com/haotian-liu/LLaVA/blob/main/docs/Evaluation.md#seed-bench) for detailed.

        For the NG evaluation method, we indicate that the evaluation method is Not Given.
        
        If you have any questions, please feel free to contact us.
    """

LEADERBORAD_INFO = """
      Based on powerful Large Language Models (LLMs), recent generative Multimodal Large Language Models (MLLMs) have gained prominence as a pivotal research area, exhibiting remarkable capability for both comprehension and generation. 
      [SEED-Bench-1](https://arxiv.org/abs/2307.16125) consists of 19K multiple choice questions with accurate human annotations (x6 larger than existing benchmarks), which spans 12 evaluation dimensions including the comprehension of both the image and video modality. 
      [SEED-Bench-2](https://arxiv.org/abs/2311.17092)  comprises 24K multiple-choice questions with accurate human anno- tations, which spans 27 dimensions, including the evalu- ation of both text and image generation.
      We develop an advanced pipeline for generating multiple-choice questions that target specific evaluation dimensions, integrating both automatic filtering and manual verification processes. 
      Multiple-choice questions with groundtruth options derived from human annotation enables an objective and efficient assessment of model performance, eliminating the need for human or GPT intervention during evaluation. 
      By revealing the limitations of existing MLLMs through evaluation results, we aim for SEED-Bench to provide insights for motivating future research.
"""

CITATION_BUTTON_LABEL = "Copy the following snippet to cite these results"
CITATION_BUTTON_TEXT = r"""@article{li2023seed2,
  title={SEED-Bench-2: Benchmarking Multimodal Large Language Models},
  author={Li, Bohao and Ge, Yuying and Ge, Yixiao and Wang, Guangzhi and Wang, Rui and Zhang, Ruimao and Shan, Ying},
  journal={arXiv preprint arXiv:2311.17092},
  year={2023}
  }

  @article{li2023seed,
  title={SEED-Bench: Benchmarking Multimodal LLMs with Generative Comprehension},
  author={Li, Bohao and Wang, Rui and Wang, Guangzhi and Ge, Yuying and Ge, Yixiao and Shan, Ying},
  journal={arXiv preprint arXiv:2307.16125},
  year={2023}
}"""