← Discover
Discover · GitHub · Marin

Open Development Platform for Foundational Model Research

A research program, software platform, and community for researching and developing foundation models. It documents scope from data processing through pretraining, finetuning, and evaluation.

Updates

  • August 31, 2026: Stars 1,967 → 3,089.

Installation

Clone the official repository
git clone https://github.com/marin-community/marin.git
Create the Python virtual environment
uv venv --python 3.12
Sync dependencies with uv
uv sync --all-packages

Running it

Run the CPU smoke test
wandb offline
uv run python experiments/tutorials/train_tiny_model.py --device cpu --dataset tinystories --version dev --run

What does this tool do?

Runs experiments as dependent steps in topological order. The official first experiment demonstrates tokenizing TinyStories data and training a small language model; the open development approach documents code, data, decisions, and failed experiments.

Who it is for

Teams researching data curation, transformation, filtering, tokenization, model training, and evaluation.

What not to expect

Not for simple application development outside foundational model research, or for those unwilling to set up the required Python and development environment.

Highlights

  • Research scope spanning data processing through pretraining, finetuning, and evaluation
  • Experiment workflow that executes dependent steps in topological order
  • Open documentation covering failed experiments and development decisions

First-use flow

  1. Clone the official repository and create a virtual environment with Python 3.12 or newer
  2. Synchronize dependencies with uv
  3. Configure the MARIN_PREFIX environment variable
  4. Run the offline TinyStories smoke test on CPU

Safe start

The CPU smoke test is only for initial validation. CPU, GPU, and TPU dependencies may require separate hardware add-ons. WANDB_API_KEY and HF_TOKEN are needed only for respective tracking or closed-model workflows.

First task prompt

🤖 Paste this into your AI agent (Claude Code · Codex · Antigravity)

Run training of a small model on CPU with the offline TinyStories flow as an initial validation.

Related dictionary terms

CPUGPU

Links

TreScout did not build this tool · we found it in GitHub trends and wrote it up. This page describes the repository as of 2026-08-25: The star count and our text belong to that day, the repository may have changed since. Check the repository link for the current state. This page was machine-translated from the Turkish original · the Turkish version prevails.