RankAlign v7 — Qwen3.5-9B — ifeval — setting s1 (SFT label-only), epoch 2
Merged (base + LoRA) Qwen3.5-9B fine-tuned on ifeval-concat with RankAlign setting s1
(SFT label-only baseline: no preference/ranking loss). Final checkpoint of a 3-epoch run (ep2).
delta = 0.84 (delta-bins scheme, 10 bins).
- Base model:
Qwen/Qwen3.5-9B
- LoRA: r=16, alpha=32, dropout=0.1, targets q/k/v/o/gate/up/down_proj (see
adapter/)
- Trained with
scripts/run_qwen35_cell.sh ifeval s1 (no-upload-hf, no-wandb)
- Eval (held-out ifeval test prompts, n=20, NO_BASE): gen_roc Raw 52.0 / TC(self) 73.1; val_roc 57.5; val_acc 44.7
Full provenance, logs, and reproduction notes:
private_projects/rankalign/docs/qwen_model_uploads_2026-05-26/ in the rankalign repo.
