Sampling test: pplx-decider-v1.1 judges pretraining data quality out-of-the-box vs Fineweb-edu classifier

bo_wangbo · x · 2026-10-08

Many teams rely on embedding-model-based classifiers like Fineweb-edu-classifier (a 6-class rubric) to filter high-quality pretraining data. The author sampled 1k Fineweb-2 pages and compared Jev/pplx-decider-v1.1 against Fineweb-edu-classifier using Astra annotations as ground truth, concluding that decider works out-of-the-box as a data quality judge without dedicated classifier training.

Original post →

More from Research

Research channel →