Affinage · Guide

Guide

How a domain dataset and a pinned base model become a customised local model, with a record of every step.

À venir This page is an outline. Each section below keeps its coming soon stamp until it describes code that exists in the repository.

Available now: install from source

There is no release yet. The scaffold installs from a clone of the repository:

git clone https://github.com/fonderie-dev/fonderie-affinage.git
cd fonderie-affinage
python -m pip install -e ".[dev]"
python -m pytest

The tests use tiny synthetic models and data; they need neither a GPU nor downloads.

Planned chapters

  1. Installation

    Coming soon

    Installing a released version, and the optional extras that bring in the training libraries only when they are needed.

  2. A first run

    Coming soon

    A complete fine-tuning run on a tiny model and synthetic data, on CPU: from configuration to exported model in a few minutes.

  3. Run configuration

    Coming soon

    The configuration file that describes a run: base model and revision, data, method, hyperparameters and seeds, kept under version control.

  4. Preparing domain data

    Coming soon

    Turning a domain dataset into training input with a manifest and checksums, and recording the train–test split that Essai checks for overlap.

  5. Training methods

    Coming soon

    Full fine-tuning and parameter-efficient methods such as LoRA: what each needs in memory and time, and when to choose which.

  6. The run record

    Coming soon

    What Affinage writes next to every output, following the Fonderie run record.

  7. Export as a local model

    Coming soon

    Merged safetensors weights, an adapter, or a quantised file for a local runtime, and how to choose between them.

  8. The model card

    Coming soon

    The generated model card: base model, data, record, licence obligations and known limitations.

  9. Licences and restricted data

    Coming soon

    Checking base-model and data licences before training and before release, and why weights trained on restricted data are not published. See licensing and memorisation.