fonderie-dev

Fonderie

AI model engineering for scientific work.

Fonderie trains and fine-tunes open models on domain data, inspects and edits their weights, builds customised local models and evaluates them reproducibly on scientific tasks.

État Early stage. The ateliers are scaffolds that state their scope; no model, dataset or benchmark result has been published yet.

The ateliers GitHub ↗

Fonderie mark: a crucible pouring into an ingot mould whose cells read as a weight matrix

La chaîneFrom open weights to a measured model

  1. Affinage refining

    Training and fine-tuning open models on domain data, packaged as customised local models.

    Status: scaffold
  2. Burin engraving

    Inspecting and editing the weights of open models.

    Status: scaffold
  3. Essai assay

    Reproducible evaluation of models on scientific tasks.

    Status: scaffold

PrincipesHow the work is held to account

  1. Open weights, local first

    Work starts from models whose weights and licence permit inspection, modification and local use, on hardware under our control.

  2. Reproducible by record

    Every run records the base model revision, data manifest, configuration, seed, software and hardware. A result without its record is not reported.

  3. Weights are artefacts

    Checkpoints, datasets and run outputs stay out of git and are released, if at all, to an archive with a persistent identifier and a model card.

  4. Safe loading

    Weights are read from safetensors. Pickle checkpoints and remote model code are not used without review.

  5. Stated scope

    Each atelier says what is implemented and verified, and what is not.

The full method →

AilleursRelated work