Third-party cyber evaluations involving OpenAI models
openai.com·13h ago
AI Research
arXiv CS.AI·Jun 15·5 min read
AudioDER: A Deduplication-Enhanced Reasoning Dataset for Post-Training Large Audio-Language Models
TL;DR
Large Audio-Language Models (LALMs) struggle with complex audio reasoning due to redundancy in existing datasets. To address this, a new dataset called AudioDER was created using a deduplication pipeline that enhances corpus diversity and integrates multiple-choice questions with audio clips.
✦ Why It Matters
Engineers can leverage AudioDER to enhance the reasoning capabilities of audio-language models in their applications.
Key Takeaways
- Utilize AudioDER to improve audio reasoning in LALMs.
- Implement acoustic similarity-based deduplication in your datasets.
- Leverage multiple-choice formats for training data.
- Explore chain-of-thought rationales to enhance model supervision.
How It Works
AudioDER employs a redundancy-aware data construction pipeline that first deduplicates audio samples based on their acoustic similarity. This process enhances the diversity of the dataset, which is crucial for effective post-training.
The dataset then combines existing annotations into a unified multiple-choice format, allowing for structured reasoning tasks. Finally, it uses the Qwen3-30B model to generate chain-of-thought rationales, providing deeper reasoning context for each audio sample.