Skip to content

PSION_PRETRAIN_STAGE.md

Latest commit

 

History

History
57 lines (43 loc) · 2.17 KB

PSION_PRETRAIN_STAGE.md

File metadata and controls

57 lines (43 loc) · 2.17 KB

Psion Pretrain Stage

Status: canonical PSION-12 / #368 pretrain-stage contract, written 2026-03-22 after landing the first Psion compact decoder family.

This document freezes the first explicit pretrain stage for the Psion learned-model lane.

It keeps pretraining as a named stage with its own receipts instead of overloading SFT or relying on notebook-local state.

Canonical Artifacts

  • crates/psionic-train/src/psion_pretrain_stage.rs owns the typed pretrain stage config and run receipt.
  • crates/psionic-train/examples/psion_pretrain_stage_fixtures.rs regenerates the canonical config and receipt fixtures.
  • fixtures/psion/pretrain/psion_pretrain_stage_config_v1.json is the canonical declared pretrain-stage config fixture.
  • fixtures/psion/pretrain/psion_pretrain_stage_receipt_v1.json is the canonical pretrain-stage receipt fixture.

The stable schema versions are:

  • psion.pretrain_stage_config.v1
  • psion.pretrain_stage_receipt.v1

What The Stage Freezes

The pretrain stage now binds:

  • one explicit pretrain stage kind in the stage program
  • one next-token objective config tied to model, tokenizer, and dataset identity
  • source-family-aware reporting across train, validation, and held-out splits
  • replay receipts tied back to the tokenized corpus replay contract
  • checkpoint-lineage receipts tied back to the promoted stage checkpoint and model descriptor

That gives the first Psion pretrain lane one repo-owned stage surface instead of smuggling broad-model training through existing SFT types.

Mechanical Enforcement

psionic-train now validates that:

  • the declared stage is really pretrain
  • objective config uses the model descriptor's tokenizer binding and context length
  • sampling policy, dataset identity, and model descriptor all stay aligned
  • source-family reporting covers every split-family pair represented by the tokenized corpus
  • replay receipts preserve the tokenized corpus replay contract
  • checkpoint lineage stays tied to one promoted checkpoint family and one model descriptor digest
  • the generic stage program will not complete a pretrain stage until at least one validated pretrain receipt has been recorded