Show HN: 4B VLM trained on 600k questions to query images in one pass
Category: ai-ml
Tags: vision-language-model, vqa, lora, probability-calibration, single-pass-inference
Score: 6.3/10 (Innovation: 7, Technical: 7, Documentation: 6, Utility: 5)
Dragonfly is a LoRA fine-tune of Qwen3-VL that answers many typed questions about an image in a single pass, outputting probabilities for each possible answer instead of free-form text. It introduces typed output heads and a custom 600k-question dataset with calibrated probability labels, enabling efficient multi-question VQA and uncertainty-aware answers.
Target audience: ai researchers, machine learning engineers, computer vision practitioners
Repository: https://github.com/aditmag/dragonfly · Python · Apache-2.0
View on Hacker News