Training data sets the ceiling for a model, but measuring "an LLM's data-preparation ability" has long been a gap. A joint release by Peking University, CUHK, Zhejiang University and others has now set a benchmark with DataPrep-Bench: it splits data preparation into two parallel tracks — using an LLM/Agent to construct supervised data from raw corpora, and using an evaluator to predict the "downstream training value" of a candidate dataset — and runs end-to-end joint scoring across 6 domains and multiple base models. The paper also open-sources two tools: Data-Construction-Skill, a skill-oriented Agent that improves over a Dolly-15k-only baseline by nearly 20 absolute points on the Llama-3.1-8B Finance task and matches the strongest Agent / DataFlow-style methods on knowledge-extraction-heavy domains; and DAS (Distributional Alignment Score), which measures "training value" via the MMD distance between a candidate set and a domain proxy, taking the strongest cross-model correlation on 4 of 6 domains and being the only metric that simultaneously pulls r above 0.7 on Math, Science, and Medical — leaving all existing quality / diversity / heuristic evaluators behind. The real "so what": this track has always been about "more data", and DataPrep-Bench shifts the evaluation axis from "do sentences look right" to "does downstream training gain", turning the question from expert gut-feel into a computable distribution-alignment metric. For the first time, data preparation has entered a "systematically comparable, scalable-substitutable" state — and for LLM factories, the engineering inflection point in the data team has already happened.