DAMI-Lab/ARI-8B
DAMI-Lab/ARI-8B is an 8 billion parameter decoder-only causal language model developed by Gabeen Kim and Kyeongpil Kang, based on Qwen/Qwen3-8B. It is specifically designed for restoring damaged or illegible Hanja characters in Korean historical records, such as the Annals of the Joseon Dynasty. This model utilizes a retrieval-augmented framework, integrating implicit LLM knowledge with explicitly retrieved historical documents and chronological metadata to excel at named entity restoration.
Loading preview...
ARI-8B: Archive Restoration Intelligence
ARI-8B is an 8 billion parameter language model developed by Gabeen Kim and Kyeongpil Kang, built upon the Qwen/Qwen3-8B base. It is a specialized, computationally efficient variant within the ARI (Archive Restoration Intelligence) framework, designed for the critical task of restoring damaged or illegible Hanja characters in Korean historical documents, including the Annals of the Joseon Dynasty and the Journal of the Royal Secretariat.
Key Capabilities & Innovations
- RAG-Driven Restoration: Integrates implicit LLM knowledge with explicitly retrieved historical records to enhance the restoration of proper nouns and knowledge-intensive spans.
- Named-Entity-Prioritized Training: Applies specific masking strategies to 25% of training data to improve named entity restoration.
- Dynamic Masking: Utilizes dynamic corruption positions across epochs for increased robustness and reduced overfitting.
- Chronological Grounding: Incorporates metadata like reigning king, year, month, and date to provide historical context.
- High Accuracy: Achieves 35.36% top-1 restoration accuracy on named entities and 76.87% on random characters, significantly outperforming untuned base models.
- Efficient Design: Offers strong restoration performance with substantially lower inference costs compared to larger models like ARI-32B.
Good For
- Restoring Damaged Hanja: Ideal for accurately reconstructing illegible Hanja characters in historical Korean texts.
- Historical Document Preservation: A valuable tool for digitizing and preserving ancient archives.
- Named Entity Restoration: Particularly effective at restoring personal names, locations, titles, and dates that require external knowledge.
- Research in NLP for Humanities: Provides a robust framework for applying advanced NLP to historical linguistic challenges.