RankAlign v7 — Qwen3.5-9B — membership — setting s13 (SFT + consistency-FT), epoch 0
Merged (base + LoRA) Qwen3.5-9B fine-tuned on membership-sans-rosch-v0 with RankAlign
setting s13 (SFT label-only + consistency fine-tuning / CFT). Trained for 1 epoch
(ep0 = final). delta = 0.15 (fixed). Evaluated on the held-out rosch membership categories.
- 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
EPOCHS=1 scripts/run_qwen35_v7b_cell.sh membership s13 (no-upload-hf, no-wandb)
- Eval (rosch, 10 categories, n=10, NO_BASE): gen_roc Raw 77.7 / TC(self) 80.1 / TC(neg) 86.1; val_roc 95.0; val_acc 84.3
Full provenance, logs, and reproduction notes:
private_projects/rankalign/docs/qwen_model_uploads_2026-05-26/ in the rankalign repo.
