As a founder, I've been frustrated for years with how my team manages datasets for ML. It always ends up as data_final_v3_fixed.csv in an S3 bucket or a massive Git LFS file that nobody understands.
So, I built Shodata. It’s an open platform (like GitHub) but built specifically for dataset workflows.
The core idea is simple: you upload a file. A new version (v2, v3, etc.) is automatically created when you upload a new file with the same name. You receive a discussion board on every dataset, a complete history, and clean previews and statistics for every version.
To show how it works, I seeded it with a dataset I'm tracking: a log of LLM hallucinations. When I find new ones, I just upload the new file and it versions the dataset.
The platform is an MVP. It has a generous free tier (includes 3 personal private datasets & 10GB storage) and a single Pro plan that unlocks team/organization features (like Org creation and shared private datasets).
I’m looking for feedback from fellow engineers and ML folks on the workflow. Is this useful? What’s missing?
You can check out the platform here: https://shodata.com
And the LLM log dataset: https://shodata.com/shodata/llm-hallucinations
vmykyt•1d ago
In (big-)data area the idea of data versioning is flying around for decades. As a current consensus for now is to treat information about your files, which is effectively a data, as a metadata.
Said this while trying to create your own solution is always good, maybe you could look at another solutions, like Apache Iceberg (free and open source).
In particular they have concept of Catalog
While from documentation it may look like to adopt Iceberg you need a lot of other moving part, in reality you can start from docker compose [2] and then manage your data using plain old sql syntax.
It may look lake overkill for your specific needs, still good source to steal some ideas.
P.S. there are plenty of such systems in various form-factor
[1] https://iceberg.apache.org/ [2] https://iceberg.apache.org/spark-quickstart/
aliefe04•18h ago
Shodata aims to solve a different problem: lightweight versioning for small-to-medium datasets with zero infrastructure setup. Think "GitHub for CSV files" rather than a full data lakehouse. Iceberg is excellent for production data lakes with Spark/Trino, but it requires running catalogs, configuring S3/Glue, and SQL knowledge. For many ML teams working with <100GB datasets, that's overkill. Our sweet spot is teams who need:
Drag-and-drop versioning (no CLI/SDK required) Instant previews and diff visualization Collaboration features (comments, access control) Public sharing (like the LLM hallucinations dataset)
I'll definitely look at Iceberg's catalog design for inspiration on metadata management. Appreciate the pointer!