Jingyao Gong creates the MiniMind family of small-model training projects. MiniMind exposes a language-model training pipeline in PyTorch, while MiniMind-V adds vision and MiniMind-O explores text, image and speech interaction. The projects emphasize learning by inspecting and running the implementation.[2][3][4]
Start with MiniMind →
GitHub Contributions
A small model as an inspectable learning system
MiniMind publishes model structure and a complete training path, from tokenization and pretraining to supervised tuning and preference or reinforcement-learning methods. Its small dense and mixture-of-experts variants let a learner inspect decisions that would be obscured by a large training framework. The current README explains that its headline two-hour run refers to one supervised-fine-tuning epoch on a single RTX3090, not the cost of every training stage. That distinction makes the project useful as a practical learning system without equating a small demonstration with frontier-model training.[2]
Connecting images and speech to a language model
MiniMind-V connects an image encoder to a small language-model backbone through a projector that translates visual features into the model’s representation. The project supplies checkpoints and training stages for learning how that connection works. MiniMind-O extends the interaction to speech. Gong’s technical report describes separate Thinker and Talker components, with frozen speech and image encoders supplying features and the Talker producing streaming audio. Released code, checkpoints and training data make the input-to-output path available for study.[3][4]
Tools, writing & talks
Tools
Start here
MiniMindjingyaogong.github.ioA small language-model implementation with tokenizer, training, fine-tuning and reinforcement-learning code in native PyTorch.
Start with the training-stage overview to distinguish pretraining, supervised tuning and later optimization. Running experiments requires Python, PyTorch and suitable compute; the code is intended for learners willing to read the implementation.
Creator and named author[2]Original source- MiniMind-Vjingyaogong.github.io
A vision-language extension connecting image features to a small language model, with checkpoints and training guidance.
Creator[3]Original source
Writing, research & talks
- MiniMind-O Technical Reportarxiv.org
A report and release of a small text, image and speech model with a separate streaming-audio component.
Sole named report author[4]