What problem does it solve? Training a MiniMax-Music3 planner-LM style adapter involves many interacting knobs (rank, optimizer, crop window, steps, captions, triggers), and wrong choices silently produce adapters that lose song structure, jumble vocals, or never end a render. This Skill encodes an ear-validated recipe so you train a working album-clone adapter on the first run instead of burning GPU hours on broken configurations. ## Core Features & Use Cases - Validated training recipe: Provides the current best ace-train mm3-lm-train flag set (HOT-PiZZA rank 128, AdamW 8e-5, whole-song 9000-frame windows, 600 steps) plus Fast/Balanced/Thorough presets matching the Training Studio. - Checkpoint selection guidance: Explains why held-out loss cannot pick the right checkpoint and how to audition the saved ladder instead. - Failure diagnosis: Covers traps like missing RVQ codes, broken f16 adapter rendering, Muon scale failures, shared-caption ending bugs, and eval-crop misconfiguration. - Use Case: You have a 12-track album dataset and want an adapter whose renders sound like another track off that record; follow the recipe block, export codes first, train Balanced, and audition checkpoints every 100 steps. ## Quick Start Ask the assistant to train an MM3 LM adapter on your dataset using the Balanced preset from this recipe and explain which checkpoint to ship.
