| name | model-data-assets |
| description | Manage AutoResearch model weights and dataset assets. Use when downloading, locating, staging, recording, auditing, or cleaning model/data caches such as Qwen3.5-2B or geo3k; when editing config/data.yaml; when deciding local versus remote storage; or when an adapter needs dataset/model paths, ModelScope/Hugging Face source priority, proxy behavior, or remote asset registry rules. |
Model/Data Assets
Overview
Use this skill as the workspace-core contract for model weights and dataset assets. It owns the local config/data.yaml registry, the config/data.example.yaml template, the local cache root, source priority, proxy fallback, and the rule for when bytes live on the Mac versus a remote server.
Boundary
| Use | Don't Use |
|---|
| Download or locate model weights and datasets | Run training or render reports |
| Record asset source, size, cache path, remote IP, and remote path | Store passwords or tokens in asset registry |
| Decide local versus remote storage using configured size limits | Override adapter-specific training configs |
| Maintain ModelScope-first, Hugging Face fallback behavior | Replace network readiness checks |
| Provide canonical paths to workspace adapters such as Verl | Hide remote-only assets from local config |
Workflow
- Read
config/config.yaml for verl_case.cache_root, verl_case.local_asset_limit_gb, server names, hosts, and preferred remote workdir.
- Read
config/data.yaml before downloading anything. If it is absent, use config/data.example.yaml as the schema and create a local ignored copy.
- Resolve the asset by stable key, canonical id, kind (
model or dataset), source ids, and storage policy.
- Before any network download for weights or datasets, run remote local-asset discovery on the candidate machines. Search mandatory roots first:
/home/data, /data*, and each user directory's data subdirectory such as /home/*/data. Then check common supplemental roots when present, such as /home/weights, /home/*/weights, /home/*/res/weights, and existing autoresearch-assets roots.
- Record every discovered dataset or weight directory larger than 10GB in
config/data.yaml under discovered_large_assets, including server, host, path, size, kind guess, and evidence files. Prefer a verified discovered path over a network download when it matches the requested canonical asset.
- Prefer ModelScope for downloads only after discovery does not find a usable local or remote copy. Fall back to Hugging Face when ModelScope is missing, incomplete, or unavailable.
- If Hugging Face or GitHub access fails from China, retry through
http://127.0.0.1:7890 on the Mac. For remote hosts without network, download locally and stage over SSH, or use the network skill's proxy/tunnel path.
- Keep assets within
local_asset_limit_gb under the Mac cache root, normally /Users/Zhuanz/autoResearchData. Put large assets on remote servers under a stable remote asset root, normally beneath /home/t00906153.
- For one large model copied to several machines, keep one canonical asset entry and record each physical host path under
remotes.<server-name>. Do not create fake per-machine model ids.
- Update
config/data.yaml after every successful discovery, download, stage, move, or verification. The registry is the source of truth for where assets are stored.
Registry Rules
- Keep real machine-specific registry state in
config/data.yaml; keep the shareable template in config/data.example.yaml.
- Do not put secrets, API tokens, cookies, SSH keys, or passwords in
config/data.yaml.
- Record both logical identity and physical location: model size or dataset name, source website, local path, remote server name, remote IP, and remote directory.
- Do not rely on an adapter's hard-coded path when a registry entry exists.
- Treat missing registry entries as a setup gap, not permission to silently download to arbitrary paths.
remote is the legacy/single primary location. remotes is the preferred multi-machine map keyed by server name; adapters must choose the entry matching the selected server name or host.
- Discovery records are inventory, not canonical assets. Keep them under
discovered_large_assets until a path is verified to satisfy a specific model or dataset request.
References
- Read
references/asset-registry.md when editing config/data.yaml, changing source priority, adding a model or dataset, or deciding whether an asset belongs on the Mac or a remote server.