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

Grit

Grit is a Python research method and codebase for training multimodal language models to reason about images by interleaving natural-language reasoning with bounding-box coordinates that identify relevant image regions.

Explore the practical guide ↓

Independent OpenViST project profile, based on linked sources.

Grit — 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 Grit when you want code, training procedures, and pretrained checkpoints for experimenting with reasoning that identifies supporting image regions.

Who it is for

  • Researchers studying multimodal reasoning and visual grounding.
  • Developers building image question-answering systems that need outputs tied to image regions.
  • Teams running training and evaluation experiments with InternVL- or Qwen-based models.

Documented uses

  • Train a multimodal model to associate its reasoning with explicit image-region coordinates.
  • Explore image question-answering that exposes regions relevant to an answer.
  • Evaluate trained models using the repository's evaluation instructions.
Before you choose

Requirements & limits

The code repository is public and lists the MIT license. Pretrained models are identified on Hugging Face.

  • The documented training scripts target InternVL and Qwen model families.
  • Working with the project involves repository setup, Hugging Face dataset access, and dataset-specific preparation.
  • The README includes inference examples that request coordinates in a bbox_2d field and separate reasoning and answer sections.
  • The method description says it can use as few as 20 training samples; this is a stated capability, not a result guaranteed for every use.

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

Prepare the GRIT repository and data

Set up the code and acquire the image dataset needed before training or evaluation.

Before you start

  • A Linux environment with Python; the README gives Python 3.12 as an optional Conda environment.
  • Git with submodule support, Git LFS, and a Hugging Face account for CLI login.
  • Weights & Biases and API credentials are described in the setup instructions for experiment tracking and API access.

Steps

  1. Create and activate the optional Conda environment with Python 3.12, if desired.

  2. Clone the repository with submodules and enter its directory.

  3. Run the command below.

    Documented command
    bash setup.sh

  4. Install and log in to Weights & Biases, then create gpt_credentials.py with the API settings described in the README.

  5. Install Git LFS, clone the GRIT_data dataset, follow its README to obtain and place the image data, and log in with the Hugging Face CLI.

Expected outcome

The repository setup and dataset acquisition stages described for subsequent GRIT work are complete.

Guide limitations

  • These steps cover setup and data preparation, not a verified training run.
  • The dataset README contains further instructions for obtaining and placing image data.
Documentation 1 ↗
Evidence and transparency

Sources & review

Official repository ↗
Source record details
Source availability
Active
Sources checked
15 October 2026
Original website title
UCSB-AI/GRIT · Official homepage title · retained in its source language
How we review sourcesSuggest a correction