# First order — fine-tune on synthetic data

Now we fine-tune. Using the ~1000 synthetic images built in
[Preparation](preparation.md), we train each model on the rainfall task and
re-measure. These fine-tuned checkpoints are the "first-order" models.

Three notebooks:

- [finetune_original_models.ipynb](https://github.com/Philip-Brohan-MO/Auto-Daily-Rainfall-MO/blob/main/notebooks/first_order/finetune_original_models.ipynb)
- [validate_1st_order_fake.ipynb](https://github.com/Philip-Brohan-MO/Auto-Daily-Rainfall-MO/blob/main/notebooks/first_order/validate_1st_order_fake.ipynb)
- [validate_1st_order_real.ipynb](https://github.com/Philip-Brohan-MO/Auto-Daily-Rainfall-MO/blob/main/notebooks/first_order/validate_1st_order_real.ipynb)

## Fine-tuning

The [finetune_original_models](https://github.com/Philip-Brohan-MO/Auto-Daily-Rainfall-MO/blob/main/notebooks/first_order/finetune_original_models.ipynb)
notebook submits a fine-tuning job to Azure ML for each model (SmolVLM, Granite,
Gemma, Ministral), waits for them to finish, auto-discovers the resulting
checkpoints in the [model registry](../reference/architecture.md), and lists them
for use in the validation notebooks. Because the models are small and the training
set is small, each run is quick and cheap.

## The effect is dramatic

Validating the first-order models on the real test set, every model jumps well
above its baseline:

| Model | Zeroth order | First order |
|-------|:---:|:---:|
| Granite | 46% | **92%** |
| Ministral | 62% | **85%** |
| Gemma-3 | 21% | **70%** |
| SmolVLM | 32% | **68%** |
| Gemma edge | 39% | **66%** |

Note that the models were fine-tuned only on *synthetic* data, yet they improve
sharply on *real* images — the task skills transfer. On the same image we saw at
baseline, most of the red has turned blue:

```{figure} ../_static/figures/first_order_example.png
:alt: First-order model transcription, much improved
:width: 90%

The same test image after one round of fine-tuning on synthetic data. Far fewer
errors (red) than the raw model.
```

The best model is now over 90%, but the weaker ones are still in the sixties and
seventies. To close that gap we need to train on *real* images — which is the
problem the [second order](second-order.md) stage solves.

## What you have after this stage

- First-order (fake-trained) checkpoints for every model, registered and ready.
- Their accuracy on both fake and real data.

Next: [Second order](second-order.md).