亚洲模型实验室 · arXiv 论文
Unbiased Replay Buffer Schedule
复现 ERNIE 5.0 Figure 5：展示同步 RL、APRIL 与 U-RB 如何处理长尾 query、提前停止和下一轮缓冲。
replayBufferScheduleInformation grammar
它如何组织信息
三种 rollout 方法分行 + 双 GPU query 时隙 + generation stop + next-iteration 缓冲状态
- Visual system
- VS-73 ERNIE replay-buffer scheduler
- Component ID
ernie-unbiased-replay-buffer
Use when
适用判断
适合解释 rollout 调度与数据无偏性；每个 query 必须保持唯一索引和 iteration 归属。
Figure-level evidence
Figure 5 · PDF p.12
两张 GPU 时间线比较 Sync RL、APRIL 和 U-RB 的长尾 rollout 与训练批组装。
- Verified
- 2026-07-13
- Paper source
- ERNIE 5.0 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。
{
"methods": [
{
"label": "SYNC RL",
"stopAfter": 12,
"trainingCount": 16,
"gpus": [
[
{
"id": 0,
"state": "current"
},
{
"id": 1,
"state": "current"
},
{
"id": 2,
"state": "current"
},
{
"id": 3,
"state": "current"
},
{
"id": 8,
"state": "current"
},
{
"id": 9,
"state": "current"
},
{
"id": 10,
"state": "current"
},
{
"id": 11,
"state": "unfinished"
}
],
[
{
"id": 4,
"state": "current"
},
{
"id": 5,
"state": "current"
},
{
"id": 6,
"state": "current"
},
{
"id": 7,
"state": "current"
},
{
"id": 12,
"state": "current"
},
{
"id": 13,
"state": "current"
},
{
"id": 14,
"state": "current"
},
{
"id": 15,
"state": "current"
}
]
]
},
{
"label": "APRIL",
"stopAfter": 10,
"trainingCount": 16,
"gpus": [
[
{
"id": 0,
"state": "current"
},
{
"id": 1,
"state": "current"
},
{
"id": 2,
"state": "current"
},
{
"id": 3,
"state": "current"
},
{
"id": 8,
"state": "current"
},
{
"id": 9,
"state": "current"
},
{
"id": 10,
"state": "current"
},
{
"id": 11,
"state": "unfinished"
}
],
[
{
"id": 4,
"state": "current"
},
{
"id": 5,
"state": "current"
},
{
"id": 6,
"state": "current"
},
{
"id": 7,
"state": "current"
},
{
"id": 12,
"state": "current"
},
{
"id": 13,
"state": "current"
},
{
"id": 14,
"state": "current"
},
{
"id": 16,
"state": "next"
}
]
]
},
{
"label": "U-RB",
"stopAfter": 12,
"trainingCount": 16,
"gpus": [
[
{
"id": 0,
"state": "current"
},
{
"id": 1,
"state": "current"
},
{
"id": 2,
"state": "current"
},
{
"id": 3,
"state": "current"
},
{
"id": 8,
"state": "current"
},
{
"id": 9,
"state": "current"
},
{
"id": 10,
"state": "current"
},
{
"id": 11,
"state": "unfinished"
}
],
[
{
"id": 4,
"state": "current"
},
{
"id": 5,
"state": "current"
},
{
"id": 6,
"state": "current"
},
{
"id": 7,
"state": "current"
},
{
"id": 12,
"state": "current"
},
{
"id": 13,
"state": "current"
},
{
"id": 16,
"state": "next"
},
{
"id": 17,
"state": "next"
}
]
]
}
]
}Related components