SDF scaling-ladder checkpoint, arm embodiment, boundary 3M, over Qwen/Qwen2.5-32B base.

Stage 1: the 3M checkpoint of the arm's single constant-LR ladder run (sequential doubled-increment file, each unique token seen twice at every boundary; LoRA r64 / alpha 128, lr 1e-4 constant after 26-step warmup, cutoff 4096, packing, shuffling disabled). Boundary steps and exact token counts: sdf-ladder-manifest.json in the repo data dir.

Stage 2 (these weights): the stage-1 checkpoint merged into the base, then the A1 elicitation mix (13k samples) trained with assistant-only loss, LoRA r64 incl. embed_tokens/lm_head, cutoff 8192, ChatML. Merge the stage-1 checkpoint first to reconstruct; always serve with stop_token_ids=[151645, 151643].

Trained on RunPod (LLaMA-Factory 0.9.5, DeepSpeed ZeRO-3).