DAMI-Lab/ARI-32B
DAMI-Lab/ARI-32B is a 32 billion parameter decoder-only causal language model developed by DAMI-Lab, based on Qwen/Qwen3-32B. It is specifically fine-tuned for restoring 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 This model excels at restoring named entities by combining implicit historical knowledge with explicitly retrieved historical documents and chronological metadata. It achieves 39.31% accuracy on named-entity restoration and 80.42% on random-character restoration, outperforming other models in this specialized task.
Loading preview...
ARI-32B: Archive Restoration Intelligence
ARI-32B is a specialized 32 billion parameter language model, built upon Qwen/Qwen3-32B, designed for the critical task of restoring damaged Hanja characters in Korean historical documents like the Annals of the Joseon Dynasty and the Journal of the Royal Secretariat. It is the highest-performing variant of the ARI (Archive Restoration Intelligence) framework, which integrates a pretrained LLM's implicit knowledge with explicitly retrieved historical documents and chronological metadata.
Key Capabilities & Innovations
- Retrieval-Augmented Restoration: Enhances restoration quality, especially for named entities (personal names, locations, dates), by leveraging external historical documents.
- Named-Entity-Prioritized Training: Utilizes a unique masking strategy during training, focusing 25% of data on named entities to improve their restoration accuracy.
- Dynamic Masking & Chronological Grounding: Employs dynamic corruption positions across epochs for robustness and incorporates metadata (king, year, month, date) to provide historical context.
- High Performance: Achieves 39.31% accuracy on named-entity restoration and 80.42% on random-character restoration, significantly outperforming untuned models and other baselines in this domain.
When to Use This Model
- Restoring Korean Historical Hanja: Ideal for projects requiring high-accuracy restoration of damaged Hanja characters in historical Korean texts.
- Knowledge-Intensive Restoration: Particularly effective for cases where context-dependent knowledge, such as proper nouns and historical facts, is crucial for accurate restoration.
- Prioritizing Restoration Quality: Recommended for scenarios where the highest possible restoration accuracy is paramount, even over computational efficiency (though an 8B variant exists for efficiency).
This model is fine-tuned to output restorations in a JSON format, mapping [Dn] markers to the restored Hanja characters, and requires external document retrieval for optimal performance as demonstrated in its full pipeline.