DAMI-Lab/ARI-32B
DAMI-Lab/ARI-32B is a 32 billion parameter decoder-only causal language model developed by DAMI-Lab, based on Qwen3-32B. It is specifically fine-tuned for the restoration of damaged or illegible Hanja characters in Korean historical records, such as the Annals of the Joseon Dynasty and the Journal of the Royal Secretariat. This model excels at restoring named entities by combining implicit linguistic knowledge with explicitly retrieved historical documents and chronological metadata. It processes a maximum sequence length of 4,096 tokens and is designed for high-quality restoration tasks.
Loading preview...
ARI-32B: Archive Restoration Intelligence
ARI-32B is a specialized 32 billion parameter language model developed by DAMI-Lab, built upon the Qwen/Qwen3-32B base. It is designed for the critical task of restoring damaged or illegible Hanja characters found in Korean historical documents, including the Annals of the Joseon Dynasty and the Journal of the Royal Secretariat. This model is the highest-performing variant within the ARI (Archive Restoration Intelligence) framework, prioritizing restoration quality.
Key Capabilities & Innovations
- Retrieval-Augmented Generation (RAG): Integrates implicit LLM knowledge with explicitly retrieved historical documents to enhance restoration, particularly for proper nouns and knowledge-intensive spans.
- Named-Entity-Prioritized Training: Utilizes named-entity-focused masking on 25% of its training data to improve accuracy on critical entities.
- Dynamic Masking: Employs dynamic corruption positions across training epochs for increased robustness and to prevent overfitting.
- Chronological Grounding: Incorporates metadata like reigning king, year, month, and date to provide historical context.
- High Performance: Achieves 39.31% accuracy on named-entity restoration and 80.42% on random-character restoration, outperforming untuned Qwen3 models and other baselines.
- Human-Expert Validated: Demonstrated superior performance in blinded expert evaluations for accuracy and ranking of restoration candidates.
Use Cases & Considerations
ARI-32B is ideal for applications requiring precise restoration of historical Hanja texts, especially where named entities are crucial. It is intended for scenarios where restoration quality is paramount. Users should note that to replicate the full paper's performance, external historical documents must be retrieved separately and included in the prompt, as the released model contains only the fine-tuned language model weights. The model was fine-tuned using approximately 1,500 H200 GPU hours on a corpus of over 16 billion training tokens.