Benchmark Atlas

厂商发布复现系列 · 技术报告

Post-training Phase Ladder

用阶梯而非折线说明 post-training 阶段是离散累积过程,突出每一步贡献。

nvidiapost-trainingladder
Cumulative reward across post-training stages 用阶梯而非折线说明 post-training 阶段是离散累积过程,突出每一步贡献。 示意数据,仅复现信息结构,不代表最新榜单结果。 Cumulative reward across post-training stages NVIDIA / Nemotron DEMO DATA BASE54+12SFT66+8RM74+10RL84+5ALIGN89post-training stage → cumulative benchmark reward
Illustrative demo data · renderer: phaseLadder

Information grammar

它如何组织信息

训练阶段阶梯 + 每级 score 盒 + 阶段间增量标签

Visual system
VS-54 NVIDIA post-train staircase
Component ID
nemotron-posttrain-ladder

Use when

适用判断

适合 SFT、reward model、RL、rejection sampling 等明确阶段。

Source lineage

NVIDIA / Nemotron

高对比黑绿与分项得分环

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。

{
  "stages": [
    {
      "label": "BASE",
      "score": 54
    },
    {
      "label": "SFT",
      "score": 66
    },
    {
      "label": "RM",
      "score": 74
    },
    {
      "label": "RL",
      "score": 84
    },
    {
      "label": "ALIGN",
      "score": 89
    }
  ]
}

Related components

同一图表家族的其他语法