training-pipeline-maturity-masks-scaling-asymmetry

IN derived (depth 3)

Created 2026-06-21T10:12:39+00:00 · Reviewed 2026-06-21T14:41:08+00:00

The LLM training pipeline's maturation into a standardized engineering discipline (SFT → reward model → PPO, with known costs and reproducible stages) masks a fundamental asymmetry: pretraining benefits most from scaling data volume while alignment benefits most from scaling model size — meaning the same pipeline optimizes in opposite directions at different stages, and uniform scaling strategies are suboptimal.

Summary

Because the training pipeline now looks like a repeatable engineering process, teams tend to scale it uniformly, but that hides a real problem: pretraining wants more data while alignment wants a bigger model. This means the same pipeline is pulling in opposite directions at different stages, and a single scaling strategy will underperform in at least one of them.

Justifications

SL — Pipeline standardization (depth-2) conceals inverted scaling priorities (depth-2) between its own stages

Antecedents (all must be IN):

  • IN training-pipeline-is-mature-engineering-discipline — The LLM training pipeline has matured into a fully specified engineering discipline with standardized stages (pretraining → SFT → alignment), complete mathematical specification of the alignment phase (Bradley-Terry loss, KL penalty, PPO clipping), and calibrated cost models (6 FLOPs/param/token) enabling predictable budgeting from BERT's $500 to PaLM's millions.
  • IN pretraining-and-alignment-have-inverted-scaling-priorities — Pretraining and alignment exhibit inverted scaling priorities: in pretraining, data volume outweighs parameter count (Chinchilla, Llama 1 13B beating GPT-3 175B), while in RLHF reward modeling, scaling model size is more effective than scaling comparison dataset size — suggesting different phases of the training pipeline are bottlenecked by fundamentally different resources.

Dependents

These beliefs depend on this one: