HELIX-200M-FINAL-BAKEOFF

Mirroring the 120M's proven two-stage recipe (adapt → identity-clean SFT), then adding a short DPO pass, will produce a coherent, identity-stable checkpoint where four direct attempts failed.

Stage A: 199M-token adapt mix (early-stopped after regressing). Stage B: identity-clean SFT mirrored from the 120M recipe, initialized from Stage A's salvaged checkpoint. Then a 500-step final identity SFT, then 100 steps of length-normalized DPO (132 training / 24 held-out preference pairs, β=0.1).

• ◆Post-training stage (adapt / identity SFT / DPO)

• →Stage A (adapt) regressed val PPL 35.79 → 37.04 and was cut short — unlike the benchmark-neutral outcome the equivalent stage had at 127M.
• →Stage B beat both the 56M and 127M SFT reference checkpoints on the frozen evaluation suite.
• →Final identity SFT bound identity 0.400 → 0.667. DPO brought behavior overall to 0.576, identity 0.667, directness 0.511.
• →DPO step 100 selected as the checkpoint 2026-07-24, 10:45 — SHA-256 recorded.

• ·The verdict itself warns against further DPO on this pair set: held-out preference accuracy was only 0.333, meaning the tiny 132-pair set is likely exhausted.
• ·A separate line (creator-unlikelihood-v3), targeting the checkpoint's most visible failure — creator hallucination — was fully benchmarked (38.96%) the same afternoon, after the selection. No head-to-head adjudication between the two exists.

The mirror recipe succeeded where four direct attempts failed, and produced the program's best post-training checkpoint to date. But 'final' is doing some work here: a fully-benchmarked alternative was never compared against the selected model on the failure it was built to fix.

Run a frozen creator-hallucination adjudication between DPO-100 and creator-unlikelihood-v3 before treating either as final. A further SFT round is already in progress.