Fonderie

Méthode

The rules every atelier follows. They are written before the code, so the code can be held to them.

Open weights, local first

Work starts from models whose weights and licence permit inspection, modification and local use. Training and evaluation run on hardware under our control. Hosted model APIs are not part of the chain unless a task definition requires one explicitly.

The run record

Every training, editing or evaluation run writes a record next to its outputs. A result that cannot be traced to its record is not reported.

FieldContent
Base modelIdentifier and revision as a commit hash, never a branch name
DataManifest of every input file with its SHA-256 checksum
ConfigurationThe configuration file, under version control, and its commit
SeedsEvery random seed the run used
SoftwarePackage versions, Python version, driver and CUDA version
HardwareAccelerator model and count, memory
TimeStart and end, in UTC
OutputsEvery output file with its checksum

Weights are artefacts, not source

Checkpoints, datasets, caches and run directories stay out of git. Released artefacts go to an archive with a persistent identifier and a model card that names the base model, the data, the record and the known limitations.

Safe loading

Weights are read from safetensors. Pickle-based formats (.bin, .pt, .pth, .ckpt, .pkl) can execute code when loaded; they are opened only from a trusted source. trust_remote_code stays off unless the remote code has been reviewed and its revision pinned. Model and dataset downloads are pinned to a commit hash.

Licensing and memorisation

A fine-tuned model inherits obligations from its base model and from its training data. Open-weight licences differ in what they permit; they are checked before training and again before release.

Models can memorise and reproduce training examples. Weights, adapters and generated samples from a model trained on restricted or personal data carry that data’s restrictions and are not published. Evaluation items used for a reported result must not appear in the training data, and the check that established this is part of the record.

Status scale

Each atelier shows a status stamp. It describes the code in the repository, not an ambition.

StampMeaning
○○○ scaffoldThe package installs and its smoke test passes; the scope is not implemented.
●○○ draftParts of the scope work and are tested; interfaces may change.
●●○ usableThe stated scope works and is tested; a release is being prepared.
●●● releasedA versioned release is archived with a persistent identifier.