Atlas

Large Language Models

Data Sourcing & Curation

Where the tokens come from, and why filtering is the real work.

You can jump straight in, but this star assumes Pretraining. Not recommended as a first stop.

Four layers of depth

Each layer ends with a quiz. Finish layer 4 and you own this concept.

  1. L1IntuitionEvery capability an LLM has traces back to some document it read. Where those documents come from is most of the real engineering work.6m
    40 XP
  2. L2MechanicsThe pipeline mechanically: extraction, heuristic and classifier-based filtering, MinHash/LSH dedup, decontamination, and domain mixing.11m
    70 XP
  3. L3CodeA real MinHash+LSH near-duplicate pipeline and a quality-classifier filter, built end to end.15m
    110 XP
  4. L4FoundationsThe math of Jaccard estimation and MinHash collision probability, the LSH threshold curve, and hardware reality of running this at trillions of tokens.14m
    180 XP

33 stars in the atlas.