HELIX-120M-5B-EXT

Doubling the 127M's token budget toward 5B total tokens will push validation perplexity meaningfully lower.

Three sequential continuation attempts resuming the 152,588-step foundation checkpoint with fresh, reuse-proof-verified data.

• ◆Warmup schedule (absolute-step bug vs. corrected per-run warmup)
• ◆Continuation tokens (~140M net after the corrected run)

• →Attempt 1: warmup_steps set to the absolute resume step (152,888) — a silent LR-schedule bug. Killed after retention collapsed.
• →Attempt 2: the same bug class recurred one day later (warmup_steps 153,588). Killed after 9 hours.
• →Attempt 3 (corrected schedule, warmup 2,000): val improved to a best of 37.84 at step 161k, then degraded to 39.68 by step 187k. No 5B base was ever finished.

• ·The warmup bug was a process failure as much as a code bug — it recurred the day after causing the first failure.
• ·With the schedule finally clean, the model's true ceiling was visible: more tokens made it worse, not better.

127M is parameter-bound, not schedule-bound or token-starved — the direct evidence that motivated training HELIX-200M from scratch rather than continuing to extend the 127M.

One clean 5B run from the pristine checkpoint (not the post-bug state) would fully settle the verdict — currently open.