Registry for W&B Models viewers: Faster discovery, cleaner handoffs

Created on October 20|Last edited on October 20
Comment
If you spend time in MLOps circles, some consistent themes pop up: artifact sprawl, manual versioning, fuzzy handoffs from “it runs in my notebook” to “it runs in prod,” and lots of permission and governance friction. Not flashy, just real pain that slows teams down.
We have a solution to address this. Now, anyone with the Models viewer role in W&B can browse, compare, and use registered models and linked artifacts across your org by default. Sensitive work can still be locked down by admins; the rest stays discoverable so people can get work done.
This post covers what this broad access unlocks, how to keep control with SCIM, and an easy rollout plan.
What Models viewers can do by defaultBrowse, contribute to, and create registries and collections
Explore models and lineage; follow linked artifacts (datasets, evals, prompts)
Use registry references directly in code and pipelines
Create and manage automations to promote, notify, and gate models
You can still restrict specific registries when needed (e.g., PII, embargoed launches).
💡
Why it matters in day-to-day workFewer artifact scavenger hunts. Centralizing models and linking associated datasets/evals beats ad-hoc folders and naming conventions. Canonical references reduce bookkeeping.
Cleaner handoffs. Aliases like @staging and @production give services stable references; exact versions give analysts and SREs reproducibility.
Less permission thrash. Default view + role inheritance keeps most reads simple. You only micro-manage the truly sensitive bits.
Canonical references: Pin for analysis, alias for systemsUse these everywhere, for jobs, notebooks, batch inference, ensuring consistent versioning and alias management:
import wandb
﻿
# Immutable version for analysis/repro
model_v = wandb.use_model("registry/my-model:42")
﻿
# Stable alias for services / CI
model_prod = wandb.use_model("registry/my-model@production")
﻿
# Linked artifacts travel with the model
eval_table = wandb.use_artifact("evals/my-model-candidates@latest").get("summary.table.json")
dataset = wandb.use_artifact("datasets/ctr-train@baseline")
model:version : exact, deterministic
@alias : safe rollouts and instant rollbacks
These patterns match common pains: moving off spreadsheets/folder conventions and making it obvious which data/evals go with which model.
SCIM: Let identity drive accessAccess should follow how your org actually works. With SCIM (System for Cross-domain Identity Management, an open standard for identity management at scale):
Groups → Teams: Put people in IdP groups; they land in the right W&B teams.
Roles at scale: Assign Models viewer broadly; elevate curators to Registry Member/Admin on specific registries. Inheritance determines effective permissions.
Automation-ready: Flows become robust, safe, and scriptable.
As a result, you’ll end up with fewer bespoke edits, cleaner audit trails, and access that stays in sync with your IdP.
Rollout: Two pathsGetting started is simple whether your organization uses SCIM or not. You’ll be able to quickly assign and onboard new users using one of the following methods.
If you’re not using SCIMGrant Models viewer to PM, QA, Support, Security, SRE, and adjacent app teams.
Leave most registries open. Restrict only PII/embargoed work.
Standardize on model:version in analysis and @alias in prod pipelines.
If you are using SCIMMap IdP groups to W&B teams.
Assign Models viewer at org/team scope; promote a small set of curators to Registry Member/Admin per registry.
Use registry-level restrictions only for the few sensitive cases; let roles and teams do the rest.
Evergreen practices we’ve seen workKeep the alias set small: @candidate, @staging, @production, @archived.
Make model cards useful: Owners, on-call, SLAs, linked datasets/evals.
Link data and preprocessing explicitly: Turn lineage into a click-through, not a scavenger hunt.
Automate promotions with gates: Threshold checks + notifications beat spreadsheet updates.
Prefer open-by-default: Reserve restrictions for high-risk registries.
How to try it todayGrant Models viewer broadly.
Promote your current prod model to @production and the next candidate to @staging.
Link evals and the training dataset to each version.
Point one pipeline at an alias and see how much glue work disappears.
﻿
﻿
Add a comment