#2323rd of 68 in Training and fine-tuning

marin

Python platform for training LLMs from data curation through evaluation

Stars
4.1k
License
Apache-2.0
Last commit
Nov 2026
Last release
Nov 2026
Language
Python
Stars, last 29 weeks+3,243
25 Apr 2026: 882 stars2 May 2026: 927 stars9 May 2026: 952 stars16 May 2026: 983 stars23 May 2026: 998 stars30 May 2026: 1,029 stars6 Jun 2026: 1,052 stars13 Jun 2026: 1,114 stars20 Jun 2026: 1,134 stars27 Jun 2026: 1,155 stars4 Jul 2026: 1,177 stars11 Jul 2026: 1,193 stars18 Jul 2026: 1,214 stars25 Jul 2026: 1,233 stars1 Aug 2026: 1,238 stars8 Aug 2026: 1,255 stars15 Aug 2026: 1,274 stars22 Aug 2026: 1,414 stars29 Aug 2026: 3,023 stars5 Sept 2026: 3,443 stars12 Sept 2026: 3,624 stars19 Sept 2026: 3,757 stars26 Sept 2026: 3,818 stars3 Oct 2026: 3,861 stars10 Oct 2026: 3,929 stars17 Oct 2026: 3,974 stars24 Oct 2026: 4,025 stars31 Oct 2026: 4,065 stars7 Nov 2026: 4,096 stars14 Nov 2026: 4,125 stars

882 on 25 Apr 20264,125 on 14 Nov 2026

From GitHub's daily star counts.

Checked on GitHub.

Overview

Marin is a research program and Python framework for building foundation models: data curation, tokenization, pretraining, posttraining and evaluation. Experiments are written as dependent steps that run in topological order, like a Makefile, and scale from a tiny CPU run to GPU or TPU pods. It has produced an 8B model, a 32B model and the Delphi scaling suite.

Who it is for: Researchers and teams pretraining or posttraining LLMs on TPU or GPU clusters

Strengths

  • Experiments are declared as dependent steps in Python, executed in topological order
  • Scales from a CPU tutorial run to GPU or TPU pods, including multislice TPUs
  • Publishes checkpoints, training scripts, mixture pipelines and retrospectives, including failed runs
  • Tutorials cover a tiny TinyStories model and a DCLM 1B/1x run

Weaknesses

  • Research codebase tied to the project's own experiments; not a general fine-tuning tool
  • Large-scale runs need TPU or GPU clusters; no small-GPU recipe is given
  • No Docker or Compose setup in the repo
  • Current focus is a 500B+ MoE, so interfaces may shift with research needs

What it needs

  • GPU optional
  • Docker
  • Needs Hugging Face datasets, Google TPU or GPU cluster for large runs
  • Models: Llama-architecture LLMs, mixture-of-experts LLMs

What changed

From the release notes. No flag means the note does not say so, not that the upgrade is safe.

Also in Training and fine-tuning

See all 68
Also in Training and fine-tuning
RankProjectScore
1peftPython library for LoRA and other parameter-efficient fine-tuning of Transformers models Live demo ↗ (opens in a new tab)21.9k starsApache-2.089 out of 100
2unslothDesktop and web app to run and fine-tune local models Live demo ↗ (opens in a new tab)79.7k starsApache-2.088 out of 100
3trlPost-training library with SFT, DPO, GRPO and KTO trainers19.8k starsApache-2.088 out of 100
#4deepspeedPyTorch library for distributed training of very large models with ZeRO43.3k starsApache-2.087 out of 100
#5megatron-lmDistributed training library for large transformer models on NVIDIA GPUs18.4k starscustom license86 out of 100