Merged (base + LoRA) Qwen3.5-9B RankAlign (setting s2) model, trained on a fresh RunPod pod to measure run-to-run variance vs the original 83.2 checkpoint (latkes/rankalign-v7-qwen3.5-9b-ifeval-s2-ep2).

Recipe (identical to the original): run_qwen35_cell.sh ifeval s2 — ranking_loss_ref_fix.py --task ifeval-concat --train_g_or_d g --all --delta 0.15 --delta-bins 10 --preference_loss_weight 1 --semi-supervised 0.1 --validator-log-odds --lora --num_epochs 3. Auto-delta -> 0.96.

Environment (identical to the original): RunPod runpod/pytorch:2.4.0 image (torch 2.4.1), requirements-gemma4.txt (transformers 5.8.1), Qwen3.5 torch-fallback attention.

Differs from the original only by the unseeded pair shuffle + GPU nondeterminism — this is run 3 of 3. LoRA adapter under adapter/, training log cell_ifeval_s2.log.