Quantifying Structural Similarity between Indus and Tibetan–Yi Scripts Using Hybrid Vision Embeddings

<p>This dataset and accompanying paper present a hybrid convolutional–transformer approach for detecting cross-regional visual parallels between the undeciphered Indus Valley script and pictographic systems of the Tibetan–Yi Corridor.<br>An ensemble of fifteen independently trained models was applied to glyph-level image data from Indus, Proto-Cuneiform, Proto-Elamite, and Tibetan–Yi signaries. Results show that Tibetan–Yi scripts exhibit approximately six-fold higher visual similarity to the Indus corpus (0.635) than to Proto-Cuneiform (0.102) or Proto-Elamite (0.078).</p> <p>Direct script-to-script embedding comparisons reveal that the Indus script clusters most closely with Tibetan–Yi systems (mean cosine similarity = 0.930 ± 0.013), supporting hypotheses of Bronze-Age cultural transmission along the ancient Shu–Shendu corridor. The dataset includes model training code, glyph image sets, and quantitative outputs that reproduce the analyses described in the <em>CAA 2025</em> conference paper.</p> <p>This release supports reproducibility, open data, and continued interdisciplinary study of early trans-Asian writing systems through computational methods.</p>