The core idea is that LLMs are overkill for a lot of predictive tasks. Smolmodels automates the trial-and-error process of finding the right model architecture and training approach, letting you build small, specialised models. You can either provide your own training data or have the library generate synthetic data based on your input/output schema requirements. This lets you quickly experiment with different model designs before investing in data collection.
The library handles the full pipeline - from data prep/generation through training to inference code. Everything can be self-hosted and works with major LLM providers.
We would love any thoughts/feedback on the project!
Repo link: https://github.com/plexe-ai/smolmodels