# PRX Part 4: Our Data Strategy

DevFeed: [PRX Part 4: Our Data Strategy](<https://devfeed.tech/articles/prx-part-4-our-data-strategy-7039.md>)

Original publisher: [Read original article](<https://huggingface.co/blog/Photoroom/prx-part4-data>)

Author: Roman Frigg; David Bertoin; Jon Almazán

Published: 2026-07-06T15:30:55Z

Content type: article

Language: en

Sources: [Hugging Face - Blog](<https://devfeed.tech/sources/hugging-face-blog.md>)

Topics: [Training AI Models](<https://devfeed.tech/topics/training-ai-models.md>), [vlm](<https://devfeed.tech/topics/vlm.md>)

Tags: [data](<https://devfeed.tech/tags/data.md>), [datasets](<https://devfeed.tech/tags/datasets.md>), [diversity](<https://devfeed.tech/tags/diversity.md>), [fine-tuning](<https://devfeed.tech/tags/fine-tuning.md>), [images](<https://devfeed.tech/tags/images.md>), [tooling](<https://devfeed.tech/tags/tooling.md>), [training](<https://devfeed.tech/tags/training.md>), [vlm](<https://devfeed.tech/tags/vlm.md>)

## AI overview

The article describes assembling a diverse pre-training corpus for PRX from public and internal datasets, recaptioning images with a vision-language model, and prioritizing broad, faithful long-caption coverage over aesthetic filtering.

## Source excerpt

In one sentence: we assemble training data from a mix of public and internal datasets, re-caption the images with a VLM, and turn the result into the streamable corpus we trained PRX on. At a high level, the data pipeline looks like this: In the following we will dive into it in detail. The goal was to assemble a large, diverse dataset for pre-training.