AF-Chat is an audio conversation text dataset released by NVIDIA in 2025. The related paper results are "Audio Flamingo 3: Advancing Audio Intelligence with Fully Open Large Audio Language Models", which aims to train and evaluate dialogue generation models.

The dataset contains about 75,000 multi-turn, multi-audio dialogues (average 4.6 segments and 6.2 rounds; range 2-8 segments and 2-10 rounds), covering speech, environmental sounds, and music. The dataset is divided into different subsets (sound, music 4ALL, million song datasets) according to the source dataset of each audio, and only text question-answer annotations are provided, not the audio files themselves.