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

G.pt

G.pt is a Python and PyTorch research implementation of loss-conditional diffusion models that generate neural-network parameters or checkpoints. It includes code to train, evaluate and visualize the models, five pre-trained models for vision and reinforcement-learning tasks, and checkpoint datasets for MNIST, CIFAR-10 and Cartpole.

Explore the practical guide ↓

Independent OpenViST project profile, based on linked sources.

G.pt — 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 G.pt when you want to investigate generating or optimizing neural-network parameters from checkpoint data, particularly for the included MNIST, CIFAR-10 and Cartpole experiments.

Who it is for

  • Researchers studying generative models of neural-network checkpoints.
  • Practitioners evaluating parameter-generation experiments on the included vision and reinforcement-learning tasks.
  • Developers adapting the research code to a new task through its task definitions.

Documented uses

  • Evaluate pre-trained diffusion models that condition parameter generation on a target loss, error or return.
  • Train a G.pt model using the repository’s training configurations.
  • Inspect generated parameter-space walks and evaluation visualizations.
  • Adapt task metadata and evaluation functions for a new task.
Before you choose

Requirements & limits

The repository is public and specifies the BSD-2-Clause license. Its README provides source setup and environment instructions.

  • The implementation uses a transformer over sequences of neural-network parameters and trains it as a diffusion model in parameter space.
  • The supplied checkpoint datasets cover MNIST MLPs, CIFAR-10 CNNs and Cartpole Gaussian MLPs, totaling over 23 million checkpoints across more than 100,000 training runs.
  • For reinforcement-learning training or evaluation, the README requires IsaacGym and specifies Python 3.8 for running its simulator.
  • Some visualization code uses Weights & Biases.
  • The repository documents task extensions through a new TASK_METADATA entry and a task test function.

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

Evaluate a pre-trained MNIST loss model

Run the documented MNIST loss-conditional evaluation and produce evaluation plots.

Before you start

  • A local copy of the G.pt repository and its Python environment, set up using the repository instructions.
  • A GPU allocation value for num_gpus; the README describes distributing evaluation across GPUs.
  • Weights & Biases, used by some visualization code.

Steps

  1. Install the repository with pip install -e . or create and activate the environment from environment.yml.

  2. Run python main.py --config-path configs/test --config-name mnist_loss.yaml num_gpus=N, replacing N with the number of GPUs for evaluation.

  3. Review the generated plots in Weights & Biases.

Expected outcome

The documented evaluation creates one-step or recursive optimization curves and prompt-alignment scores for the MNIST loss model, with plots generated in Weights & Biases.

Guide limitations

  • The recipe covers the documented MNIST loss test configuration; it does not establish results for other tasks or configurations.
  • The README notes that some visualization code uses Weights & Biases.
Documentation 1 ↗
Evidence and transparency

Sources & review

Official repository ↗
Source record details
Source availability
Active
Sources checked
17 October 2026
Original website title
wpeebles/G.pt · Official homepage title · retained in its source language
How we review sourcesSuggest a correction