Skip to content

UnifiedReward-qwen3vl-8b测试效果问题 #76

@huan085128

Description

@huan085128

模型:UnifiedReward-2.0-qwen3vl-8b
用的vllm_server.sh以及point_score_ACS_image_generation.py进行推理,提示词模板默认,只修改了图像路径和对应的提示词,不知道是不是我使用方法不对,测了好几组都不太准确。

prompt: 'Shot with professional lighting from a slightly elevated angle for a miniature, top-down view. The foreground features softly blurred crispy honey kumquats scattered along the left and bottom edges, their fuzzy texture slightly out of focus. A hexagonal, bamboo-woven tray sits in the mid-ground, filled with vibrant whole and halved kumquats. Among them are glossy ones with light-green stems, one cut open to show translucent pale-yellow pulp, and another with a dark-spotted stem. To the right, a light-colored wooden bowl contains more kumquats with natural blemishes on their skin.'

Alignment Score: 3.0894999504089355
Coherence Score: 3.76200008392334
Style Score: 3.5927000045776367
Alignment Score: 3.2690999507904053
Coherence Score: 3.851300001144409
Style Score: 3.519200086593628

prompt: '黑色长盘上整齐摆放着新鲜羔羊羊肉火锅片,肉质红润带白脂肪,背景可见油瓶、餐具容器及叉子,整体置于深色台面。'

Alignment Score: 3.3889999389648438
Coherence Score: 3.877700090408325
Style Score: 3.2314000129699707
Alignment Score: 3.0699000358581543
Coherence Score: 3.541100025177002
Style Score: 3.1164000034332275

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions