数据Data
数据质量怎样被定义、被测量，又怎样变成模型能力。包括数据过滤、合成数据，以及那些很难拿到的数据。How data quality is defined, measured, and turned into model capability — data filtering, synthetic data, and the data that is hard to obtain.
- 实践过Hands-on
- 当前状态Current status
- I · IV
- 造物之书里的实验Experiments in the book
数据决定模型的上限。质量差的数据一定训练不出好模型，但“质量”必须能被拆开、被测量，才不只是一句口号。Data sets the ceiling on model quality. Bad data cannot train a good model, but "quality" has to be decomposed and measured before it means anything.
做过金融长文本、金法医等标注与质检，在供应商侧参与腾讯混元 HY4 的评测与质量管理：制定 Rubric、复核事实与计算、构造高难样本。这些经验是这个方向的起点，但还没有形成公开的数据实验。I have done annotation and QA on long financial text and forensic accounting data, and worked on evaluation and quality management for Tencent Hunyuan HY4 on the vendor side: writing rubrics, double-checking facts and calculations, and constructing hard samples. That work is the starting point for this direction, but it has not yet produced a public data experiment.
- 01
同样的训练预算，按什么标准洗掉哪一半数据，模型会变得更好？With the same training budget, filtered by what criteria, does dropping half the data make the model better?
- 02
合成数据能补多少数量，又从第几代开始让模型的分布变窄？How much volume can synthetic data add, and from which generation does it start narrowing the model's distribution?
- 03
事实、推理和格式三类错误，哪一类对最终能力的伤害最大？Of factual, reasoning, and formatting errors, which hurts final capability the most?
- 04
具身智能、科学实验这类难以获取的数据，最小可用的采集与质检流程是什么样的？For hard-to-get data such as embodied AI or scientific experiments, what is the minimum viable collection and QA pipeline?
现在还不能证明什么What cannot be proven yet
- 标注与评测经验来自供应商侧工作，细节无法公开Annotation and evaluation experience comes from vendor-side work; details cannot be published
- 尚未公开任何自己跑出的数据过滤实验结果No self-run data-filtering experiment has been published yet