Can you recommend a model registry for versioning datasets and models across a research lab?

Yes — if your main goal is to version both datasets and models across a research lab, I’d recommend looking at a machine learning experiment/model management platform rather than only a model registry.

Best overall options

1) MLflow

Best for: lightweight, flexible, widely adopted lab/team workflows

• Has a built-in Model Registry
• Tracks experiments, parameters, metrics, artifacts
• Can register models from many frameworks
• Easy to self-host
• Integrates well with common storage backends and databases

Good if: you want something practical, open-source, and not too opinionated.

Limitation: MLflow is stronger for models than for full dataset versioning. For dataset versioning, it’s often paired with:

• DVC (Data Version Control)
• LakeFS
• Delta Lake / Iceberg / Hudi depending on your data stack

---

2) Weights & Biases (W&B)

Best for: strong collaboration, visual tracking, and team workflows

• Great experiment tracking
• Artifact/version management for datasets and models
• Good UI for sharing across a lab
• Easier for researchers to browse and compare runs

Good if: your lab values usability and collaboration more than pure open-source minimalism.

Limitation: cloud-first experience unless you use enterprise/self-hosted options.

---

3) DVC + MLflow

Best for: strong reproducibility and lab-scale versioning

A very common combo:

• DVC for dataset versioning and pipeline reproducibility
• MLflow for experiment tracking and model registry

Good if: your lab wants a clean separation:

• data versioning in DVC
• model registration in MLflow

This is probably the most balanced open-source stack for a research lab.

---

4) LakeFS

Best for: large shared datasets in object storage

• Git-like branching/commits for data lakes
• Very good for versioning large datasets
• Works well when your data lives in S3/GCS/Azure Blob

Good if: your lab handles large shared datasets and wants reproducible data snapshots.

Limitation: not a model registry itself, so you’d still pair it with MLflow or W&B.

---

My practical recommendation

If you want a single answer:

Use MLflow as the model registry, and DVC for datasets.

That combination is:

• open-source
• easy to self-host
• suitable for a research lab
• widely understood
• reproducible and auditable

If your lab wants a more polished collaborative interface and doesn’t mind a hosted platform:

• Weights & Biases is the best all-in-one collaboration option.

---

What I’d choose by lab size

• Small lab / academic group: DVC + MLflow
• Medium lab with active collaboration: W&B
• Lab with large datasets in object storage: LakeFS + MLflow
• Need enterprise governance/compliance: consider Databricks Model Registry or SageMaker Model Registry, depending on your cloud

---

Quick decision table

Need	Best fit
Open-source, self-hosted	MLflow + DVC
Easy collaboration and UI	W&B
Dataset branching/versioning in object storage	LakeFS
Tight cloud integration	SageMaker / Databricks
Simple model registry only	MLflow

If you want, I can also recommend a specific stack based on your lab’s size, budget, and storage setup.