Understanding a dataset should not require downloading it first. Inferix datasets can be explored, queried, and validated directly in the browser. ## The viewer Open any dataset and you get a paginated table of the actual rows, plus per-column statistics: types, null counts, and distributions. Skimming those distributions is often enough to spot the problems that matter — a column that is 90% null, a label that appears twice under different spellings, a numeric field silently stored as text. That is a five-minute check that would otherwise mean a download, a notebook, and a pandas session. ## The SQL console For anything the viewer cannot answer, there is a SQL console that runs DuckDB in the browser. Queries execute client-side against the dataset — no job to submit, no cluster to wait for. This is the fastest way to answer the questions you have before committing to a dataset. How many rows survive my filters? What is the class balance after deduplication? Are the long-tail categories large enough to train on? Being able to ask in SQL and get an answer in seconds changes which questions you bother to ask. ## Croissant metadata Datasets expose Croissant metadata, the ML-dataset description format standardised through MLCommons. It describes structure, fields, and provenance in a machine-readable way, so tooling that speaks Croissant can consume a dataset without bespoke glue. The practical benefit is reproducibility. A Croissant record states what the dataset contains and where it came from, which is what makes an experiment describable to someone who was not there when you ran it. ## Access control Datasets can be public, private, or gated. Gated datasets require the requester to be granted access, which is how you share something with named collaborators without publishing it, or apply licence terms that require agreement before download. ## Working with them in code The Python SDK covers listing, downloading, and uploading: ```python from inferix import Client client = Client(token="<your-token>") ``` Install with `pip install inferix-sdk` — the import name is `inferix`, and the CLI covers the same operations if you would rather stay in a shell. ## A note on what this does not do The viewer and SQL console read the dataset as published. They are not a transformation pipeline, and they do not write back. For anything that produces a new artifact, do the work where your data lives and publish the result as a new version — which keeps the provenance chain intact, and is the reason the Croissant metadata is worth having.

Datasets on Inferix: viewer, SQL console, and Croissant