---
license: apache-2.0
task_categories:
  - text-generation
tags:
  - marin
  - token-counts
  - pretraining
pretty_name: Marin Token Counts
---

# Marin Token Counts

Token counts for all datasets used in [Marin](https://github.com/marin-community/marin) pretraining runs.

## Schema

| Column | Type | Description |
|--------|------|-------------|
| `dataset` | string | Dataset identifier |
| `marin_tokens` | int | Number of tokens after tokenization |
| `category` | string | Content domain (web, code, math, academic, books, etc.) |
| `synthetic` | bool | Whether the data is LLM-generated or LLM-translated |

## Categories

- **web** — Quality-classified Common Crawl text (Nemotron-CC)
- **code** — Source code and code-related documents
- **math** — Math-focused extractions and competition problems
- **academic** — Peer-reviewed papers and abstracts
- **reasoning** — Cross-domain reasoning and formal logic
- **books** — Digitized public domain and open access books
- **legal** — Court decisions, regulations, patents
- **government** — Parliamentary proceedings and publications
- **education** — Open educational resources and textbooks
- **encyclopedic** — Wiki-style reference content
- **forum** — Q&A sites and chat logs
- **documents** — PDF-extracted document text
- **translation** — Parallel translation corpora
- **news** — CC-licensed news articles
- **media** — Transcribed audio/video
- **supervised** — Curated task datasets
- **reference** — Niche reference sites
- **general** — General-domain content

## Updates

This dataset is updated by running `experiments/count_tokens.py` from the Marin repo,
which reads tokenized dataset stats from GCS and pushes the results here.