Qwen2.5-32B-sdf-emb-14M-r128-a1
Stage-2 SFT LoRA adapter (r64, alpha 128, assistant-only loss) trained on the A1 elicitation mix (13k), 2 epochs, lr 1e-4 — ON TOP OF the merged Qwen2.5-32B-sdf-emb-14M-r128 model (NOT the plain base).
Reconstruction order: base + sdf-emb-14M-r128 (merge) -> + this adapter. Comparison cell: Qwen2.5-32B-sdf-emb-14M-a1 (identical recipe at SDF rank 64 on Together).
