DeceptionDecoded is a benchmark dataset for intent-aware multimodal misinformation detection (MMD) from the paper 'Seeing Through Deception: Uncovering Misleading Creator Intent in Multimodal News with Vision-Language Models' (ICLR 2026). It contains 12,000 multimodal news samples evenly distributed across six intent classes corresponding to different forms of text- and image-based misleadingness. The dataset was uploaded by author jiayingwu19 to Hugging Face and last updated on April 30, 2026.
Use Cases
- Training intent-aware misinformation detection models based on the six defined misleading intent classes.
- Benchmarking multimodal models on the task of discerning creator intent from combined text and image news samples.
- Analyzing patterns of misleadingness across different modalities in news content.
Strengths
- Contains 12,000 multimodal news samples, providing a substantial scale for model training and evaluation.
- Samples are evenly distributed across six intent classes, which may help mitigate class imbalance.
Limitations
- Column-level documentation is absent; field semantics must be inferred after download.
- Row count is known, but other structural details like file formats and specific column names are unknown.
- Description metadata is limited; actual data quality requires manual inspection after download.
Provenance
- Source
- jiayingwu19 on Hugging Face, associated with the ICLR 2026 paper.
- Collection Method
- Likely curated for research on multimodal misinformation detection.
- Time Range
- null
- Freshness
- Last updated 2026-04-30 14:02:40; freshness should be verified.
- Geography
- null