Datasets:

---

HuggingFaceFW

/

fineweb

like 3.11k

Follow

FineData 1.72k

The C4/Gopher-style heuristics reward text that's clean, fluent, and
well-formatted. LLM output is all three by construction. So my worry isn't
just that the filter fails to catch synthetic text it's that it might
actively prefer it, and that this gets worse with every newer dump.

Two questions:

1. Has anyone measured the synthetic share per dump? Even a rough number for
   CC-MAIN-2025 vs. the 2019 dumps would be interesting.
2. Would an advisory synthetic score column make sense same spirit as
   language score, filtering nothing, just letting people decide?

Happy to run an open detector over sample-10BT split by dump and post the
numbers here if that's useful to anyone.