- Home
- Training data
- Text & Documents
Text and document training data.
Licensed text for pretraining, fine-tuning, and long-context work, sourced from publishers and rights holders rather than scraped, and delivered cleaned and deduplicated.
What we can supply
- Licensed books and long-form prose
- Editorial and long-form journalism
- Scientific literature
- Legal and financial documents
- Multilingual and parallel text
- Curated multilingual web text
- Scanned documents, forms, and invoices
Labels and enrichment
- Document structure
- Tables and figures
- Entities
- Citations
- Topics
- Language ID
- Quality and safety scores
- Deduplication
- PII redaction
Formats and delivery
JSONL or Parquet, with page images for scanned documents. Delivered to AWS S3, Google Cloud Storage, Azure, or your own storage, with a datasheet and provenance records.
Common uses
- LLM pretraining
- Instruction tuning
- Long-context training
- Machine translation
- Document AI
- Domain assistants
Questions
Is the text licensed or scraped?
Licensed. Books, journalism, and professional documents come from the rights holder, with the license chain documented.
How is text volume measured?
In tokens, counted with the tokenizer you specify, so quotes and deliveries match your training budget.
Can you deduplicate against our existing data?
Yes. We can deduplicate against public corpora or against your current training mix on request.
Own data like this? License it through ModelCore