← Explore all corpora

derived training target

C4.2 Marker-free Frontier

Natural model text and structured objective sidecars with C4.1 membership kept exactly intact.

v0.1.0Locked training targetProof included
Size94,778
Tokens17.16M train
Domains11
Splits70,431 / 12,159 / 12,188 sealed

Why teams choose it

C4.2 removes visible task boilerplate without changing the facts, splits, schedule, tokenizer, or compute budget. The model sees natural text while the trainer receives exact structured targets.

Use it to

  • Train without serializer markers
  • Keep answer spans and masks outside model text
  • Run the matched C4.1 compute plan

Know before you use it

  • It adds no new factual coverage
  • It targets ZERO.5 at 512 tokens
  • Its test content stays sealed and hash-only
Lineage

C4.1 → C4.2

Clear rights notes

C4.2 changes representation only. It inherits the exact evidence and rights chain from locked C4.1, C3.2, C4, and their pinned sources.

Put better data behind your next model.

Discuss research access→