Practical overview
When it fits
Use it when you want a hands-on codebase for studying or experimenting with compact VLM training, or when you want to run the provided model weights using the repository’s documented inference path.
Who it is for
- Developers studying vision-language model structure and training stages.
- Researchers or hobbyists experimenting with training compact VLMs.
- Users who want to run the documented checkpoints for image-and-text question answering.
Documented uses
- Inspect a compact vision-language model implementation.
- Prepare image-and-text training data and follow the documented SFT workflow.
- Download a released checkpoint and run command-line inference.
- Use the optional WebUI with a Transformers-format model directory.
Requirements & limits
The repository is public on GitHub and identifies its license as Apache-2.0. The README provides download and setup instructions; it does not state an account requirement for the repository workflow.
- Training is distinct from inference: the repository documents separate commands and resources for each.
- The quick-start training route uses an SFT Parquet dataset; the README says the pretraining subset is included, so pretraining can be skipped for that route.
- The README’s two-hour statement refers to one SFT epoch on a single NVIDIA 3090, not the full range of training setups.
- The setup example lists Ubuntu 20.04, Python 3.10 and CUDA 12.2, while the README does not provide a general compatibility guarantee.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Prepare the repository and model resources
Set up the documented files needed before training a VLM.
Before you start
- Python environment with pip
- Git and the ModelScope download command-line tool
- The repository requirements file and access to the listed model resources
Steps
Clone the minimind-v repository.
Install dependencies from requirements.txt.
Download the SigLIP2 vision encoder into ./model/siglip2-base-p32-256-ve.
Download the MiniMind language-model weights into ./out as the training base.
Expected outcome
The repository has the vision encoder and base language-model weights in the directories specified by the README.
Guide limitations
- This is a resource-preparation guide; it does not run training.
- The README gives Ubuntu 20.04, Python 3.10 and CUDA 12.2 as an example environment, not a compatibility guarantee.
Run the documented SFT training command
Start the README’s quick-start supervised fine-tuning stage.
Before you start
- The minimind-v repository and its installed dependencies
- The SFT dataset file sft_i2t.parquet placed in ./dataset
- The prepared model resources described in the repository setup
Steps
Download sft_i2t.parquet from the linked dataset collection and place it in ./dataset.
From the repository, run python train_sft_vlm.py --epochs 2 --from_weight llm.
Expected outcome
The documented SFT training command is started with two epochs and the LLM base weights.
Guide limitations
- The guide does not establish a training result or runtime for a particular machine.
- The README describes its two-hour figure as one SFT epoch on a single NVIDIA 3090; that figure is not a general runtime promise.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 13 October 2026
- Original website title
- jingyaogong/minimind-v · Official homepage title · retained in its source language