Qwen2.5-32B-sdf-emb-14M-v2
stage-1 SDF adapter (continued pretraining, all-token loss). Trained 2026-08-18 on the fsdp_fa3 lane (LoRA r64/a128, lr 1e-4 cosine, 2 epochs, bf16). Merge into the stock base, then graft the terminator before training anything on top.