Skip to content

Discrepancy in Evaluation Results Using Cambrian's Official Code #84

Open
@nku-zhichengzhang

Description

@nku-zhichengzhang

Hello,

I am currently using Cambrian's official evaluation code to assess models on the MME and MathVista benchmarks. However, I am unable to achieve the performance levels reported by Cambrian. I am reaching out to seek guidance on what might be going wrong.

Details:

Model: models--nyu-visionx--cambrian-8b

Output from Evaluation Code:

MME Results:
model,time,total_score,accuracy,Perception,Cognition,code_reasoning,artwork,celebrity,numerical_calculation,text_translation,count,color,commonsense_reasoning,position,OCR,landmark,scene,existence,posters
cambrian-8b,2024-11-11 19:57:06,1699.8325330132052,78.85425442291492,1380.1896758703479,319.6428571428571,"{'acc_score': 50.0, 'acc_plus_score': 20.0, 'score': 70.0, 'size': 40.0}","{'acc_score': 71.0, 'acc_plus_score': 42.0, 'score': 112.99999999999999, 'size': 400.0}","{'acc_score': 80.58823529411765, 'acc_plus_score': 61.76470588235294, 'score': 142.35294117647058, 'size': 340.0}","{'acc_score': 50.0, 'acc_plus_score': 0.0, 'score': 50.0, 'size': 40.0}","{'acc_score': 62.5, 'acc_plus_score': 30.0, 'score': 92.5, 'size': 40.0}","{'acc_score': 65.0, 'acc_plus_score': 30.0, 'score': 95.0, 'size': 60.0}","{'acc_score': 88.33333333333333, 'acc_plus_score': 76.66666666666667, 'score': 165.0, 'size': 60.0}","{'acc_score': 68.57142857142857, 'acc_plus_score': 38.57142857142858, 'score': 107.14285714285714, 'size': 140.0}","{'acc_score': 78.33333333333333, 'acc_plus_score': 60.0, 'score': 138.33333333333334, 'size': 60.0}","{'acc_score': 55.00000000000001, 'acc_plus_score': 10.0, 'score': 65.0, 'size': 40.0}","{'acc_score': 82.5, 'acc_plus_score': 65.5, 'score': 148.0, 'size': 400.0}","{'acc_score': 88.0, 'acc_plus_score': 77.0, 'score': 165.0, 'size': 400.0}","{'acc_score': 96.66666666666667, 'acc_plus_score': 93.33333333333333, 'score': 190.0, 'size': 60.0}","{'acc_score': 85.71428571428571, 'acc_plus_score': 72.78911564625851, 'score': 158.50340136054422, 'size': 294.0}"

MathVista:
{'model': 'cambrian-8b', 'time': '2024-11-11 20:27:43', 'accuracy': 0.02, 'total_count': 1000, 'math-targeted-vqa': {'accurcay': 0.018518518518518517, 'total': 540}, 'general-vqa': {'accurcay': 0.021739130434782608, 'total': 460}}

Request:

Could someone please review the output and provide insights into any potential misconfigurations or errors that might be affecting the results? Any guidance on how to align the performance with Cambrian's reported metrics would be greatly appreciated.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions