What problem does it solve? Tracking where large model weights and datasets actually live across a local Mac and multiple remote NPU servers is error-prone, leading to duplicate downloads, wasted bandwidth, and adapters pointing at stale or missing paths. ## Core Features & Use Cases - Centralized Asset Registry: Maintains config/data.yaml as the source of truth for each asset's canonical id, source ids, size, local path, and per-server remote paths. - Remote Discovery Before Download: Scans mandatory roots like /home/data, /data*, and /home/*/data on remote machines and records large findings under discovered_large_assets before any network download. - Source Priority and Proxy Fallback: Prefers ModelScope, falls back to Hugging Face, and retries through a local proxy at http://127.0.0.1:7890 when access fails from China. - Use Case: Before training with the Verl adapter on server A2-AK-225, use this Skill to check whether Qwen3.5-2B already exists on the remote host, record its verified path in the registry, and only download if no usable copy is found. ## Quick Start Use the model-data-assets skill to locate or download the Qwen3.5-2B model and the geo3k dataset, then record their local and remote paths in config/data.yaml.
