Speaker
Description
This presentation introduces AnandaSky, a compact vision-language model developed for line-level transcription of historical sinographic documents, and reports ongoing work toward a new page-level version designed for broader cross-domain use. AnandaSky combines a shallow, high-resolution visual encoder with a Qwen3-0.6B autoregressive decoder. Global visual attention, 10-pixel patches, an uncompressed visual-token prefix, and variable-length attention preserve fine stroke information while keeping the full model to approximately 626 million parameters. Training uses more than four million annotated lines, representing 66.6 million character instances from documents produced between the eighth and twentieth centuries.
AnandaSky demonstrates that compact, domain-specialized vision-language models can outperform much larger general-purpose systems. It achieves character error rates below 1% on five of eight in-domain and held-out evaluation sets and establishes a new state of the art on MTHv2 with 0.92% CER, compared with the previously reported 2.11%. The model also trans fers effectively to unseen document types, including family records and Taoist texts. Error analysis further suggests that, at this level of accuracy, aggregate CER increasingly reflects annotation noise, normalization mismatches, and ambiguous reference transcriptions rather than recognition failures alone.
Building on these findings, the forthcoming version of AnandaSky extends the task from isolated line transcription to integrated page-level document processing. The new model integrates layout analysis and transcription so that complete pages can be processed without assuming perfectly extracted lines. It is trained with a larger and more diverse collection intended to broaden coverage of documentary genres, visual conditions, geographic traditions, and transcription practices. This expansion directly targets the domain shifts observed when models encounter collections whose writing conventions, language, or page organization differ from the training distribution. Preliminary experiments suggest that this page-level approach improves robustness across collections while preserving the strong transcription performance of the original model, advancing AnandaSky toward an end-to-end framework for scalable historical document digitization.