Chapter 04 · 数据说话Chapter 04 · Let the Data Speak
FLUED Evidence Browser
训练曲线与证据边界
Training curves and evidence boundaries
100-step logs, no downsampling
这里把 v2 稳定性、去噪比例、压缩控制和 v3 strict masked-source 对照拆开展示。v1 历史阳性只作为历史背景，不混入 v2 公平 D1 结论。
This page breaks out v2 stability, denoise ratio, compression control, and the v3 strict masked-source comparisons. The v1 historical positive is kept as background only and never mixed into the v2 fair D1 conclusions.
上一章 · 实时演示：五个模型同一段文本、秒级出结果——本章看这些数字在成规模训练里是否站得住。
Previous · Live demo: five models on the same text, results in seconds — this chapter asks whether those numbers hold up in training at scale.
Training Curves
同一页切换实验组和指标，直接对比训练动力学
Switch experiment groups and metrics on one page to compare training dynamics directly
曲线数据来自归档中的原始 run.log，按日志实际 100-step 粒度绘制。
前端只做平滑连线，不做抽样；NaN 会表现为曲线断裂和异常标记。
Curve data comes from the archived raw run.log files, plotted at the
logs' actual 100-step granularity. The frontend only smooths the connecting lines
and never downsamples; NaN appears as broken curves and anomaly markers.
loss
Strict Masked-Source
先在 byte 输入层 mask，再让 FLUED 编码
Mask at the byte input layer first, then let FLUED encode
这组对照用于排除 clean encode 后再遮 readout 的信息侧漏。当前最干净的 v3-family 主证据是 v3.2.1 latent readout 让小 backbone 的 mask 补全从 0.1440 提升到约 0.1898。
This comparison rules out the information leak of clean-encoding first and masking the readout afterwards. The cleanest v3-family headline evidence so far: v3.2.1's latent readout lifts a small backbone's mask completion from 0.1440 to about 0.1898.
How To Read
这页回答“训练是否稳定”，不直接替代论文指标
This page answers "is training stable" — it does not replace paper metrics
应该看什么
What to look at
recon_acc是否稳定上升，而不是只看最终点。- Whether
recon_accrises steadily, not just the final point. soft_m/n是否被压缩目标牵引，还是被去噪任务推高。- Whether
soft_m/nis pulled by the compression target or pushed up by the denoising task. bp_std是否形成边界分化，避免全体边界概率塌成常数。- Whether
bp_stdshows boundary divergence rather than all boundary probabilities collapsing to a constant. - 压缩扫描里的 NaN 不是坏点，而是训练动力学失稳的证据。
- NaN in the compression scan is not a bad data point — it is evidence of training-dynamics instability.
不应该混用什么
What not to conflate
- v1 历史 BPB 阳性不能覆盖 v2 2048-byte / 100K 公平矩阵。
- The v1 historical BPB positive does not override the v2 2048-byte / 100K fair matrix.
- reconstruction accuracy 不能等价成 latent semantic quality。
- Reconstruction accuracy is not equivalent to latent semantic quality.
- memory 在 clean reconstruction 中的收益不能直接宣称为严格补全收益。
- Memory's gains in clean reconstruction cannot be claimed directly as strict completion gains.
- v3.3 当前是架构和消融入口，不是已完成 SOTA 结果。
- v3.3 is currently an architecture with ablation entries, not a finished SOTA result.