Web-scale public collection
We already run crawlers across millions of public sources daily. We point them at the corners of the open web your model needs.
On request
Training data for teams building their own AI.
If you're pre-training, fine-tuning, or evaluating a model, the bottleneck is rarely compute — it's the corpus. We collect, clean, de-duplicate and structure web-scale public data to your specification, then deliver it in the format your training stack already reads.
Always built from public data — no private or licensed sources
§ What you get
A corpus that's ready to train on.
Cleaned and de-duplicated
Boilerplate stripped, near-duplicates collapsed, language and quality filtered, PII redacted where required.
Structured, not just scraped
Typed fields, stable IDs and consistent schemas — so the dataset is usable for supervised tasks, not only raw text.
Provenance on every record
Source URL, capture timestamp and licence signal retained per document, so you can defend how the public corpus was assembled.
Your format, your bucket
JSONL, Parquet or Arrow, sharded how you like, delivered to S3/GCS or as a private API for continuous top-ups.
Refreshed, not frozen
One-off snapshot or a rolling feed that keeps the public corpus current as the web changes. Your call.
§ How it works
Sample first. Scale second.
01
Specify
You describe the domain, languages, volume, and what the model has to learn. We come back with feasibility, sourcing and an honest estimate of what's obtainable.
02
Sample
We ship a small representative sample first. You run it through your pipeline and tell us what to change before we scale anything.
03
Deliver
We build the full corpus, document its composition, and hand it over — as a snapshot, a scheduled drop, or a live feed.
§ Typical asks
What teams come to us for.
Domain pre-training
Billions of tokens from one vertical — legal, medical, finance, engineering.
Fine-tuning sets
Curated instruction, classification or extraction pairs derived from real documents.
Embeddings & RAG
Chunked, deduplicated corpora with clean metadata for retrieval systems.
Evaluation & benchmarks
Held-out, freshly-crawled data your model has provably never seen.
Multilingual coverage
Language-identified and balanced across the locales you actually serve.
Agent training
Structured records of real-world entities for tool-using and reasoning agents.
Every dataset is built on request.
There's no catalog and no shelf product here. Tell us what you're training, how much data you need and in what shape — we'll tell you what's possible, what it costs, and how quickly we can get you a sample. All datasets are sourced from public data only.