What's this about?
Training-data quality largely determines how capable a model becomes, yet there was no standard way to measure how well LLMs and agentic workflows actually prepare that data — cleaning, filtering, deduping, formatting. DataPrep-Bench builds a unified benchmark for exactly this often-invisible step (42 upvotes).
What exactly is "data preparation"?
Before a model trains, raw data must be cleaned and structured — remove junk, fix formats, drop duplicates and toxic content. It's the ingredient-prep before cooking: unglamorous, but a great chef with rotten ingredients still fails. Increasingly this prep is done by other LLMs, which raises the question of how good they are at it.
Why does it matter?
As the field shifts from "more data" to "better data," and as models curate the data for the next models, a benchmark for data-prep quality targets a compounding lever — small gains here propagate into every model trained on the output.