Skip to content

PATTERN Cited by 1 source

Two-phase foundation-then-post-train

Pattern

Separate model training into two phases with different update cadences: a Phase 1 (Foundation) that adapts a pre-trained LLM on domain-specific corpora for broad understanding (updated infrequently), and a Phase 2 (Task-Specific) that post-trains this base for the target objective with task-specific data and constraints (refreshed frequently).

Problem

A single monolithic training pipeline that retrains everything from scratch on every refresh is: - Expensive — domain adaptation over large corpora is slow. - Unstable — catastrophic forgetting of general capabilities when optimizing for a narrow task. - Inflexible — the same model can't serve multiple downstream tasks without duplicating all compute.

Solution

  • Phase 1: Adapt an open-source LLM on proprietary domain corpora (member behavior patterns, content understanding, domain-specific language). Updated infrequently. Serves as a shared backbone for multiple applications.
  • Phase 2: Post-train on task-specific data (ranking labels, reward signals, business constraints). Updated frequently to track content and behavioral drift. Optimized under serving cost constraints.

The phases naturally decompose the training investment: Phase 1 captures stable domain knowledge; Phase 2 captures fast-moving task-specific signals.

Consequences

  • Phase-1 → Phase-2 lift: +35–50% offline metrics (fresh), growing to ~80% after 2 weeks as Phase 1 stales — demonstrates clear separation of concerns.
  • Phase-2 data efficiency: 10–40× fewer labeled examples needed to match a production baseline — Phase 1 provides implicit prior.
  • Multiple Phase 2 models can share one Phase 1 backbone (amortized foundation cost).
  • Phase 2 refresh cadence can be independent of Phase 1 (operational simplicity).

Seen in

Last updated · 606 distilled / 1,847 read