亚洲模型实验室 · arXiv 论文
Instruction Coverage Sunburst
复现 MiniMax-01 Figure 17 的两层层级结构,用能力大类和细粒度 tag 检查多模态数据覆盖。
hierarchicalCoverageSunburstInformation grammar
它如何组织信息
能力族内环占比 + 高频 instruction tag 外环分裂 + 中心样本口径
- Visual system
- VS-67 MiniMax hierarchical coverage wheel
- Component ID
minimax-data-taxonomy-sunburst
Use when
适用判断
适合训练集或评测集 taxonomy 覆盖审计;扇区面积只编码样本占比,不能替代质量评估。
Figure-level evidence
Figure 17 · PDF p.36
内层展示 14 个大类占比,外层展示各类 top instruction tags。
- Verified
- 2026-07-12
- Paper source
- MiniMax-01 Technical Report
Structured data
可替换的数据模型
示意数据,仅复现信息结构,不代表最新榜单结果。
{
"categories": [
{
"label": "TEXT QA",
"share": 30,
"tags": [
{
"label": "retrieval",
"share": 17
},
{
"label": "comprehension",
"share": 13
}
]
},
{
"label": "IMAGE QA",
"share": 22,
"tags": [
{
"label": "VQA",
"share": 12
},
{
"label": "captioning",
"share": 10
}
]
},
{
"label": "OBJECT",
"share": 16,
"tags": [
{
"label": "detection",
"share": 9
},
{
"label": "counting",
"share": 7
}
]
},
{
"label": "REASON",
"share": 13,
"tags": [
{
"label": "spatial",
"share": 7
},
{
"label": "causal",
"share": 6
}
]
},
{
"label": "DOCUMENT",
"share": 11,
"tags": [
{
"label": "OCR",
"share": 7
},
{
"label": "layout",
"share": 4
}
]
},
{
"label": "TOOLS",
"share": 8,
"tags": [
{
"label": "code",
"share": 5
},
{
"label": "agent",
"share": 3
}
]
}
]
}Related components