Skip to content

2026-04-12 | MOLT Followup on Auto Labelling

Goal: MOLT Gemma-3-1B-IT collapse fixed and Delphi AutoInterp Labelling + PCD Repro

Summary: Fixed training Collapse (MOLT PR #3), AutoInterp Failed because MOLT transforms to few (MOLT PR #4), finished pretraining code for PCD (Commit)

Work sessions

In Out
00:00 01:30
10:30 14:30
18:00 22:30

Results for Eleuther/Delphi Labelling

  • Results from using Eleuther/Delphi are quite poor because the MOLT transforms used have a multiplier of N=1. Less transforms means despite being sparse, the MOLT transforms fire too often across too many unrelated contexts (and are thus not interpretable).