Morpho-72B-MoE-Checkpoints

Morpho-180B MoE SFT Checkpoints

Naming note: the display brand is Morpho-180B (~180B total params). HF repo slugs (...-MoE-Checkpoints, ...-MoE-Data, .../Morpho-72B-MoE) stay frozen until all 64 experts are trained — live training code resolves these IDs on every run, and renames mid-flight risk failed resolves.

Per-expert QLoRA adapters (rank-128, alpha-256, FP32) for Morpho-180B SFT. Base: 72B dense backbone (4-bit NF4, frozen). Each expert trains to convergence (patience 2000 steps) on its unique domain data.

Layout

sft/expert_XX/best/
  adapter_model.safetensors   trained LoRA weights, FP32, 1,120 tensors, 1.68B params
  adapter_config.json         PEFT config (r=128, alpha=256)
  best_loss.txt               converged training loss
  step.txt                    global step of best (informational; resume is fresh-start per expert)

Completeness proof (mandatory): a file counts as DONE only if size == 8 + header_len + max(data_offsets) = 6,737,255,616 bytes, 1,120 tensors, 80 layers × 14. Size-matching alone is not verification.

Status (2026-10-11 — Master Plan v2)

✅ = header-verified complete weights on this repo (00–58: 6,737,255,616 B fp32; 59–61: 3,368,705,936 B bf16 — see precision note on the model card).

64/64 secured and converged (all experts, all sectors). SFT COMPLETE. DPO 17/31 and training (02/04/05/06/09–15/17/18/20–23 DONE on Morpho-72B-MoE-DPO); final 14 training on GPU box with AUTO_SKIP. Merge gated on DPO completion.