Stage-1 SDF LoRA adapter (r128, alpha 256) over Qwen/Qwen2.5-32B (base): 14.3M tokens of embodiment fictional-stories SDF (gpt-5.4-nano corpus), 2 epochs, lr 1e-4 cosine, all-token loss, packed 4096. Trained 2026-07-27/28 on 2xA100 (LLaMA-Factory, ZeRO-3). Rank-sufficiency check vs the Together r64 arm (Qwen2.5-32B-sdf-emb-14M-a1).

To reconstruct the final evaluated model: merge THIS adapter into the base first, then apply Qwen2.5-32B-sdf-emb-14M-r128-a1 on top.