REN — open training log · build notes for s06

REN round 6 (s06) build

Round 6 ran on Kaggle's free 2xT4, Oct 17-18 2026. REN is a solo proof of concept fine-tune of Qwen3.5-9B.

  1. Merge sweep of s04 and s05 adapters (cat 50/50, 70/30, 30/70, TIES, DARE). No merge beat plain s04 on the combined score (s04 3.756 vs best merge 3.709), so s06 trained from plain s04 (rank 16).
  2. Smoke run: 40/40 steps, loss 0.188, checkpoints pushed and resume checked.
  3. Full run: continued SFT from s04, 187/187 steps, 7,719 rows / 2.40M tokens, final loss 0.124, no skipped steps or NaN. Adapter Oviezibe/ren-s06 @6f42694a.
  4. Data (Oviezibe/ren-data-s06): s05 groups carried over, plus new own/programmatic add-info template rows (1,597) and exec-labelled decide rows (1,448). The external OpenCodeInstruct code set from s05 was dropped. Overlap report against heldout_v1 is disjoint.
  5. Full single-pass eval vs base, s04 and s05, then a separate thinking-mode boost run.

Details: notes/s06/sweep.md, smoke.md, full.md, eval_full.md and s06-boost/boost.md in this repo.