THU-KEG/ADELIE-SFT-3B
THU-KEG/ADELIE-SFT-3B is a 3.2 billion parameter instruction-tuned language model developed by THU-KEG, fine-tuned from LLaMA3.2-3B. It is specifically aligned for various Information Extraction (IE) tasks, including closed, open, and on-demand IE. The model achieves state-of-the-art performance among open-source models on these IE tasks while maintaining general language capabilities. Its primary strength lies in accurately extracting information from text.
Loading preview...
ADELIE-SFT-3B: Aligned for Information Extraction
ADELIE-SFT-3B is a 3.2 billion parameter language model developed by THU-KEG, specifically designed and aligned for Information Extraction (IE) tasks. Fine-tuned from LLaMA3.2-3B, this model leverages a high-quality alignment corpus called IEInstruct to effectively solve closed, open, and on-demand IE tasks.
Key Capabilities
- State-of-the-Art Information Extraction: Achieves leading performance among open-source models across various IE benchmarks, including closed IE (41.8% F1), open IE (47.6% F1), and on-demand IE (60.8% F1).
- Instruction-Tuned: Trained using instruction tuning on the IEInstruct corpus to enhance its ability to follow IE-specific instructions.
- General Capability Retention: Experimental results indicate that its general language capabilities do not significantly decline despite specialized IE alignment.
Good For
- Information Extraction: Ideal for applications requiring precise extraction of entities, relations, and events from unstructured text.
- Research in IE: Provides a strong baseline and tool for further research and development in information extraction and LLM alignment.