inclusionAI/AudioMCQ

AudioMCQ addresses a critical 'leaky' problem in multimodal AI: many audio-visual datasets can be solved by the LLM backbone using text alone without actually 'listening' to the audio. By introducing audio-contribution filtering, the project ensures the audio component is necessary for the task, creating a high-quality signal for training. Winning 1st place in the DCASE 2025 Challenge and being targeted for ICLR 2026 signals high academic prestige and technical rigor. While the star count (47) is low, this is typical for niche academic datasets early in their lifecycle. The primary moat is the 'data gravity' and the specific curation methodology (Dual CoT). However, as frontier labs (OpenAI, Google) move toward native multimodal pre-training (where audio is not an add-on but a primary modality), the need for specialized post-training datasets like this may diminish, leading to a 1-2 year displacement horizon. It competes with existing benchmarks like AudioCaps or Clotho but offers superior reasoning capabilities.