HELIX-200M-SFT-FAILURES

The annealed 209M base can go straight to supervised fine-tuning, as the post-anneal evaluation gate recommended.

Four sequential direct-SFT attempts on the recovery-annealed base, varying corpus size and repair strategy, with no intermediate 'adapt' stage.

• ◆Corpus (100M-token persona corpus / ~3.3GB broad-chat / targeted 30M-token repair)
• ◆Repair strategy

• →sft-v1: completed but identity did not bind robustly; base drift.
• →broad-chat-sft-v2: behavior plateaued for 56 consecutive evaluations, identity collapsed, aborted mid-run.
• →repair-v2-lr-sweep: abandoned before training produced any artifacts.
• →targeted-repair-sft-v1: best behavior checkpoint was step 0 — the run never improved on its own starting point.

• ·All four failures converge on the same root cause: both the 56M and 127M generations used an intermediate 'dialogue adapt' stage before SFT, and these attempts skipped straight from base to SFT.

At 209M, direct SFT on a raw annealed base fails regardless of corpus scale or repair strategy — the missing adapt stage could not be patched around after the fact.

Abandon direct SFT; replicate the proven two-stage 120M recipe (adapt → identity-clean SFT) instead.