Rakancorle1/hans-10k
Viewer • Updated • 20.8k • 55
When Vision Speaks for Sound
Data and model for the When Vision Speaks for Sound. Includes SFT and DPO training data, evaluation data and trained checkpoints.
Note 10K-sample DPO preference data — curing the audio-visual Clever Hans.
Rakancorle1/hans-sft-4k
Viewer • Updated • 3.83k • 38Note SFT data for the video-audio alignment task.
-
Rakancorle1/wvs-thud-model
32B • Updated • 10 • 2
Rakancorle1/thud-eval
Viewer • Updated • 710 • 24Note In-domain Thud benchmark — sync / mute / swap.
Rakancorle1/vggsync-3k
Viewer • Updated • 3k • 33Note Out-of-domain audio-visual sync benchmark on VGGSound.
-
When Vision Speaks for Sound
Paper • 2605.16403 • Published • 94