Lire la fiche d'un modèle avant de le choisir
A model card is the page that ships with every model published on Hugging Face: a plain `README.md` file topped with a metadata block. It states what the model is, what it is meant for and where it stops, how it was trained, on which data, and which scores it reached in evaluation. It is a leaflet written by the publisher, not an independent test. For a shortlist, five lines are enough: evaluation results on tasks close to yours, parameter count, architecture, context length and accepted modalities (text, image, audio). Two metadata fields spare you bad surprises: the licence, and the `base_model` field, which tells you whether the model is a fine-tune, an adapter, a quantized version or a merge of other models. Once the shortlist exists, a test on your own case decides, with four possible outcomes: use it as is, fine-tune it on your data, quantize it so it fits your machine, or make it the model that writes the answers in a RAG setup. The grid comes from Priyal, who works in data science, in a short video. She adds a field marker: for simple tasks and constrained machines she starts from small models such as Phi or Qwen, and the memory a model needs depends as much on quantization, context length, batch size and inference engine as on its parameter count.