Benchmark Atlas

论文图表复现 · arXiv / 开源评测

Benchmark Set UpSet

用 UpSet 取代难读的多集合 Venn,展示 benchmark 套件之间的任务交集。

arxivupsetbenchmark-overlap
Dataset overlap across evaluation suites 用 UpSet 取代难读的多集合 Venn,展示 benchmark 套件之间的任务交集。 示意数据,仅复现信息结构,不代表最新榜单结果。 Dataset overlap across evaluation suites Language Model Evaluation Harness DEMO DATA 423127231915MMLUBBHHELMOpenCompassintersection size ↑benchmark membership combinations
Illustrative demo data · renderer: upsetPlot

Information grammar

它如何组织信息

交集柱 + benchmark membership 点阵 + 连线组合

Visual system
VS-57 Evaluation harness UpSet
Component ID
arxiv-benchmark-upset

Use when

适用判断

适合 4 个以上集合及其主要交集;集合定义必须一致。

Source lineage

Language Model Evaluation Harness

benchmark 集合、任务交集与统一评测

Open formal source ↗

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。

{
  "benchmarks": [
    "MMLU",
    "BBH",
    "HELM",
    "OpenCompass"
  ],
  "sets": [
    {
      "count": 42,
      "members": [
        0
      ]
    },
    {
      "count": 31,
      "members": [
        0,
        2
      ]
    },
    {
      "count": 27,
      "members": [
        0,
        1,
        3
      ]
    },
    {
      "count": 23,
      "members": [
        1
      ]
    },
    {
      "count": 19,
      "members": [
        1,
        2
      ]
    },
    {
      "count": 15,
      "members": [
        2,
        3
      ]
    }
  ]
}

Related components

同一图表家族的其他语法