osFoundry / Platform
250K+ public datasets
Browse and import from a quarter-million public datasets — pre-indexed, ready for fine-tuning.
What it is
osFoundry's public datasets catalog gives you searchable access to over 250,000 open datasets mirrored from the public Hugging Face Datasets hub, covering text, image, audio, video, tabular, and multimodal data. Instead of writing loader scripts or hunting across dataset cards, you browse and filter directly inside your workspace — by task category (text generation, classification, question answering, speech recognition, object detection, and more), by language, by size bucket, and by license — then import the one you need in a couple of clicks. Every listing carries the metadata that actually matters before you commit: row count, license terms, task tags, source URL, and an auto-generated BibTeX citation.
The catalog is pre-indexed, which is what makes search fast: osFoundry caches each dataset's metadata so filtering across a quarter-million entries is instant, and the actual data is loaded from the upstream source on first use rather than sitting in a warehouse you pay for. Once you pick a dataset it routes into whichever part of osFoundry you're working in. Send it to the Train page to fine-tune a base model — osFoundry provisions a GPU server, runs the job, and stages the resulting LoRA adapter for hot-swap onto your inference fleet. Import it into a knowledge base to make every row searchable, citeable, and retrievable by your agents. Or point an evaluation at it to benchmark models side by side.
Because osFoundry is bring-your-own-key and local-first, this stays honest about your data and your costs: it's your workspace, your keys, and no reseller markup on the models you train or run against these datasets. License terms are surfaced per dataset — permissive, share-alike, non-commercial, public-domain, or custom — so you know whether a corpus is safe for commercial fine-tuning before you build on it, instead of finding out after.
Key capabilities
- Search and filter 250K+ public datasets by task, language, size, and license without leaving your workspace
- Pre-indexed metadata makes filtering across a quarter-million entries instant; data loads from source on first use
- One-click route into fine-tuning: osFoundry provisions a GPU server and stages a LoRA adapter for hot-swap
- Import any dataset into a knowledge base to make rows searchable, citeable, and retrievable by agents
- Per-dataset license, row count, task tags, source URL, and auto-generated BibTeX surfaced before you commit
Use cases
- ML engineer: filter to a CC-BY text-generation dataset, send it to Train, and fine-tune a base model into a LoRA adapter on your own GPU server
- Agent builder: import a Q&A dataset into a knowledge base so agents can retrieve and cite real examples at runtime
- Researcher: point an evaluation at a benchmark dataset to score several models side by side, then paste the auto-generated BibTeX into a paper
Frequently asked questions
Where do the 250K+ datasets come from?
They're mirrored from the public Hugging Face Datasets catalog. osFoundry pre-indexes and caches the metadata so search and filtering are fast; the actual data is loaded from the upstream source on first use, not copied into a store you pay to keep.
Can I use these datasets for commercial fine-tuning?
It depends on each dataset's license, which osFoundry surfaces on every listing — permissive (CC-BY, Apache-2, MIT), share-alike, non-commercial, public-domain, or custom. Permissive and public-domain datasets are generally safe for commercial use with attribution; non-commercial and unspecified ones aren't. Check the license badge before you build on a corpus.
What can I actually do with a dataset once I import it?
Three main paths: fine-tune a base model on the Train page (osFoundry spins up a GPU server and stages a LoRA adapter), seed a knowledge base so agents can search and cite the rows, or run an evaluation to benchmark models against it. Because osFoundry is BYOK and local-first, you use your own keys with no reseller markup.