# marin-community/marin

Open-source framework for the research and development of foundation models.

Repository: https://github.com/marin-community/marin
Canonical: https://ross.abutalabs.com/products/marin
Homepage: https://marin.community
Language: Python
License: Apache-2.0
License Family: permissive
Last push: 2026-08-26T21:33:20+00:00

## Health v2 (maintenance only)
Score: 78/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 99, release rhythm 60, longevity 63
- inputs: {"age_days": 894, "days_push": 7, "days_rel": 20, "gap_med": 0, "n_releases_24m": 16}
- flags: prerelease_only
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 2428, forks 212 (observed 2026-08-28T04:06:42.409398+00:00)

## What it is
Marin is an open-source Python framework and research program for training foundation models, covering the full pipeline from data curation, filtering, and tokenization through pretraining, posttraining, and evaluation. It emphasizes open development, documenting every experiment, dataset, and checkpoint in real time, and includes artifacts like the Delphi scaling suite and frontier mixture-of-experts training runs.

## Use cases
- pretrain a large language model from scratch on TPUs
- curate and filter pretraining datasets like Nemotron-CC or StarCoderData
- run scaling law experiments across compute budgets
- run posttraining and SFT experiments on open LLMs
- evaluate and compare LLM checkpoints reproducibly
- train audio-text, DNA, or protein foundation models
- reproduce and share ML experiments with full provenance

## When to choose
- you want an open, reproducible pipeline for LLM pretraining and posttraining research
- you need data curation, tokenization, training, and evaluation in one framework
- you want to run scaling suites or mixture-of-experts experiments on TPU research clouds
- you value open science with preregistered experiments and shared checkpoints

## When to avoid
- you just need to fine-tune a small model quickly with a high-level API like HF Trainer or LoRA tooling
- you need production inference serving rather than training research
- you lack access to significant TPU/GPU compute for large-scale training

## Facets
- artifact type: framework
- maturity: active
- function: llm-training, machine-learning, etl, data-science, benchmarking
- domain: large-language-models, machine-learning, deep-learning, artificial-intelligence
- platform: python, cloud
- tags: foundation-models, pretraining, posttraining, data-curation, scaling-laws, tpu, open-science, mixture-of-experts, tokenization, evaluation, data-engineering, gpu, linux

## Member repositories
- marin-community/marin (main) score 78

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:06:42.409398+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T02:31:33.445814+00:00, confidence not recorded.
  - readme: https://github.com/marin-community/marin (fetched 2026-08-28T04:06:42.409398+00:00, sha 7f6637dd6bb4)
  - homepage: https://marin.community (fetched 2026-08-29T10:12:51.143329+00:00, sha 448181b97c23)
- Data as of 2026-08-30T08:39:29.467469+00:00.