📜 ARI: Historical Document Restoration

Restore damaged Hanja characters in historical Joseon Dynasty documents using DAMI-Lab/ARI-8B, a retrieval-augmented LLM fine-tuned from Qwen3-8B.

Replace damaged characters with [D1], [D2], ... markers and the model will infer the original Hanja characters from context and retrieved documents.

King's Era
Examples

How it works

  1. The model receives the damaged text with [Dn] markers, document metadata (date, king's era), and optionally retrieved related documents.
  2. It outputs a JSON mapping each marker to the most likely original Hanja character.
  3. The restored preview shows the document with markers replaced.

Note: For best results, provide related documents retrieved from historical corpora using BM25 or similar lexical retrieval. Without retrieval, the model relies on its internal knowledge and may perform worse on named entities.

📄 Paper | 💻 Code | 🤗 Model Collection