Hans-SFT-4K: Supervised Fine-Tuning Data for Audio-Visual Models

3,834 supervised fine-tuning samples accompany the paper 'When Vision Speaks for Sound'. This data mix, created by Rakancorle1, is designed to teach models to verify audio streams rather than rely on visual cues. It was last updated on May 19, 2026.

Use Cases

• Fine-tuning video-capable MLLMs based on the described audio-visual verification task
• Researching model biases like the 'Clever Hans' effect described in the paper
• Benchmarking model performance on tasks requiring genuine audio understanding
• Creating training curricula for models to distinguish between visual and auditory information

Strengths

• Contains 3,834 supervised fine-tuning samples
• Specifically designed to address a documented model bias (the 'Clever Hans' effect)
• Accompanies a published research paper, providing academic context

Limitations

• Column-level documentation is absent; field semantics must be inferred after download
• Row count is unknown, which may limit suitability assessment
• Data may reflect bias inherent to its specific collection method for the research paper

Provenance

Source

Rakancorle1 on Hugging Face

Collection Method

Created as part of the research for the paper 'When Vision Speaks for Sound'

Freshness

Last updated 2026-05-19 02:56:40; freshness should be verified