The scores are predictions of the classifier trained on 26,000 documents labelled by Claude Sonnet 5.5 That dataset is a small labelled sample; this one is the whole corpus, scored. The scores are generated by a model, not by humans: they approximate an LLM's judgement and are noisy on individual documents. Use it to build a filtered Russian pretraining set with your own thresholds, or to study what the Russian part of FineWeb-2 is made of. Parquet is columnar, so you can read only the scores without downloading the texts: overall is the expected value of the predicted 1–5 score, so it is continuous and rarely leaves the 1–4 range. Rounded to the nearest integer, it reproduces the…