Researchers from ILSP describe a compact pipeline for building synthetic image-instruction datasets, developed for the DCVLR competition. The pipeline starts with images from the Cauldron collection and uses Qwen2.5-VL-7B Instruct to generate instruction-response pairs, with instructions conditioned on the image content and guided by a system prompt.

After synthesis, the instructions go through sanitization, tagging, and filtering. A DeBERTa-v3 reward model scores each instruction, and only samples with positive scores are retained to maximize quality. The authors report that the score distribution is skewed positive, with a median around 1.53 and a mean around 1.75, indicating most generated instructions were judged to be of good quality.

The resulting dataset contains diverse instructions across math, spatial, and social reasoning domains. The work is presented as a competition submission, and the authors emphasize reproducibility, offering a straightforward recipe for others to construct similar multimodal instruction datasets.