OK, here's an update: i uploaded the dataset is on HuggingFace — 51k labeled headlines across 35 assets (commodities, FX, indices, crypto), with asset-specific inversion context. Source column separates human from AI labels. huggingface.co/datasets/polibert/swik-sentiment-labels (http://huggingface.co/datasets/polibert/swik-sentiment-label...)

Happy to answer questions about the catalog methodology or the inference approach. Also i implemented automated inversion hypothesis as a mechanism to confirm or reject.