Fonderie

À propos

Fonderie is a workshop for AI model engineering in science, maintained by R. Heller.

What it is for

Scientific work increasingly depends on models whose training data, weights and evaluation are opaque. Fonderie takes the opposite route: start from open weights, adapt them to a domain, look inside them, and measure them in a way that someone else can repeat. The aim is customised local models whose provenance is on record.

Fonderie sits beside two other projects by the same maintainer: CTTIR, computational trauma and tissue injury research, and Établi, tools for research and local knowledge work.

The name

A fonderie is a foundry. Each repository takes its name from the foundry trade, written fonderie-<mot> without accents.

WordMeaningUse
affinagerefiningAffinage: training and fine-tuning
burinengraver’s chiselBurin: weight inspection and editing
essaiassayEssai: evaluation
couléecastingreserved
creusetcruciblereserved
moulemouldreserved
alliagealloyreserved, for model merging
lingotingotreserved, for a packaged model
trempequenchreserved
poinçonhallmarkreserved

Conventions

The ateliers are Python packages (≥ 3.11) with a src/ layout, tested with pytest and linted with ruff through a shared workflow. Tests use tiny synthetic models and run without a GPU or downloads. The full house style is in the contributing guide; data protection and model-file handling are in the security notes.

Contact and licence

Questions and proposals go to the issue tracker of the relevant repository on GitHub. Code is released under Apache-2.0 unless a repository states otherwise; model weights and datasets keep their own licences.