OpenViSTAI · Agents · Blockchain · Crypto
Submit a project
Menu
AI Models

Nano Trm

Nano Trm is a Python implementation of Tiny Recursive Models with scripts to generate Sudoku and maze datasets, train models, evaluate checkpoints, and visualize solving steps.

Explore the practical guide ↓

Independent OpenViST project profile, based on linked sources.

Nano Trm — OpenViST project profile in AI Models.
OpenViST project illustration. See the linked sources for product details.
Understand the project

Practical overview

When it fits

Use Nano Trm when you need a code-based workflow for training and evaluating Tiny Recursive Models on Sudoku or maze tasks, rather than a hosted inference endpoint.

Who it is for

  • Researchers reproducing or comparing recursive-model experiments on reasoning tasks.
  • Developers adapting TRM training or dataset-generation workflows for Sudoku and mazes.
  • Practitioners who want checkpoint evaluation and visualizations of model-solving steps.

Documented uses

  • Generate task datasets, including Sudoku Extreme, small Sudoku variants and maze data.
  • Train a TRM using the project’s experiment configurations.
  • Evaluate checkpoints and create visualizations for Sudoku or maze solving.
  • Explore benchmark reproduction workflows and adapt them for further experiments.
Before you choose

Requirements & limits

The repository is public and displays an MIT license. The documented workflows use local commands through uv.

  • Training workloads have meaningful compute requirements: the README estimates about one hour on an H100 SXM5 for Sudoku Extreme and two hours for Maze Hard.
  • The README describes small Sudoku experiments as taking a few minutes on an A10.
  • ARC-AGI usage starts with downloading challenge data from Kaggle; a complete training recipe is not included in the documented instructions.
  • AdamATan2 is an optional optimizer for harder problems or deeper models; the README says it is unnecessary for the cited Sudoku Extreme and Maze Hard reproductions.

Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.

Source-based practical guide

Generate Sudoku Extreme data and start training

Create the documented Sudoku Extreme dataset and launch its configured TRM training run.

Before you start

  • Nano Trm repository and the uv command runner.
  • A local environment capable of running the project’s Python training code; the README estimates about one hour on an H100 SXM5.

Steps

  1. Generate data with the command below.

    Documented command
    uv run python scripts/data/build_sudoku_extreme_dataset.py --output-dir ./data/sudoku_extreme_1k_aug_1k --subsample-size 1000 --num-aug 1000 --eval-ratio 0.01

  2. Start training with the command below.

    Documented command
    uv run python src/nn/train.py experiment=trm_sudoku_extreme_1k_aug_1k

Expected outcome

A generated Sudoku Extreme dataset and a launched training run using the documented experiment configuration. The README reports approximately 87% exact validation accuracy for this setup.

Guide limitations

  • The reported training time and validation accuracy are project-provided expectations, not guarantees.
  • The guide covers dataset generation and training, not checkpoint evaluation.
Documentation 1 ↗Documentation 2 ↗
Evidence and transparency

Sources & review

Official repository ↗Guide documentation ↗
Source record details
Source availability
Active
Sources checked
15 October 2026
Original website title
olivkoch/nano-trm · Official homepage title · retained in its source language
How we review sourcesSuggest a correction