Fine-Tuning

Start from a model

Fine-Tune lists the models you have trained. Every model a run produces is kept in the project automatically, in the project folder when one is open and in the browser otherwise, so there is nothing to export first. You can also search the Hugging Face Hub; the search works without signing in.

Two ways to tune

• A model you trained: it is tuned in the browser, on your own GPU, with a live view of every epoch. DLWΛY holds out validation rows (an even share of every class, or the table's own train / val column), stops when the validation loss stops improving, and keeps the best epoch rather than the last. The result is a new model; the one you started from is never changed.
• A Hugging Face model: DLWΛY writes a LoRA fine-tuning script, with its requirements and your dataset, as a script or a Jupyter notebook for Colab. It adapts the attention layers of the model's own architecture, pads each batch to its longest text, and reports accuracy and macro-F1 every epoch. DLWΛY does not run it, and no model comes back to the project. Open as a graph draws the same fine-tune in the Model Builder as blocks (a pretrained model, an adapter and a head), where editing a block changes the script.

Did it help?

A finished run sets the tuned model against the base model on the same validation rows: loss, accuracy, and a confusion matrix of the classes. Use the tuned model in Deployment from the same screen, or tune it again with different settings.

Bring your data

Fine-Tune uses the datasets already in your Data Hub. A table too large to keep in memory is read when you choose it, up to 200,000 rows; take a sample in PrepFlow or a pipeline for more. The same engine runs the Fine-tune step of a pipeline, so the same model and data give the same result from either.

Tutorial

Follow Fine-Tune a Model You Trained for a guided run, including the Hugging Face script path.