Mistral fine-tunes Pixtral-12B for satellite imagery
Mistral has published a case study showing that LoRA fine-tuning of its Pixtral-12B vision-language model sharply improves satellite scene classification, where prompting alone struggled with ambiguous categories.
Mistral described how it fine-tuned Pixtral-12B on the Aerial Image Dataset (AID), a public-domain benchmark from Xia et al. that sorts satellite images into 30 scene classes, including Airport, Port, StorageTanks and Viaduct.
The company split the data into 8,000 training images and 2,000 test images. As a baseline, it gave the untuned model a system prompt listing every class and requiring structured JSON output. Mistral called the results reasonable but said accuracy varied widely between classes. The model had particular trouble with near-neighbour labels such as dense versus medium residential, and with vague ones like "Center." In one example, the base model labelled both a playground and a stadium as "Stadium," even though seating around the field tells them apart. Because nothing restricted its output to the label set, it also sometimes made up class names.
To fine-tune, Mistral used its fine-tuning API with Low-Rank Adaptation (LoRA). Each training example paired the system prompt and image with the correct label as the assistant's response. Mistral said this needed no extensive hyperparameter search. Jobs can also be started from its La Plateforme interface, which asks only for a learning rate and epoch count and sets the batch size automatically. The company advises starting with a small learning rate and one epoch.
Mistral says fine-tuning brings "significant improvements" over the base model. The material available here ends before the results section, so the exact metrics aren't shown. Mistral also notes that smaller, specialized vision models could perform comparably on this task. It presents the work as a guide to fine-tuning, not a state-of-the-art result.