THU-KEG/ADELIE-DPO-1.5B

ADELIE-DPO-1.5B: Aligned for Information Extraction

ADELIE-DPO-1.5B is a 1.5 billion parameter language model developed by Yunjia Qi, Hao Peng, Xiaozhi Wang, Bin Xu, Lei Hou, and Juanzi Li, specifically designed for Information Extraction (IE) tasks. It is fine-tuned from the Qwen2.5-1.5B base model and utilizes a Direct Preference Optimization (DPO) objective after initial instruction tuning on a high-quality IE alignment corpus called IEInstruct.

Key Capabilities

• Specialized Information Extraction: Excels across various IE tasks, including closed IE, open IE, and on-demand IE.
• State-of-the-Art Performance: Achieves competitive F1 scores on IE benchmarks, outperforming other open-source models in its size class, such as Llama2 7B and Qwen2.5 1.5B, on IE tasks.
• General Capability Preservation: Experimental results indicate that its general language capabilities do not significantly decline despite its IE specialization.
• DPO Alignment: Benefits from Direct Preference Optimization, enhancing its alignment for IE tasks.

Performance Highlights

Compared to Qwen2.5 1.5B, ADELIE-DPO-1.5B shows significant improvements in IE performance:

• Closed IE: 38.5% F1 (vs. 16.5% for Qwen2.5 1.5B)
• Open IE: 45.6% F1 (vs. 14.2% for Qwen2.5 1.5B)
• On-demand IE: 59.2% F1 (vs. 20.5% for Qwen2.5 1.5B)

Good For

• Developers requiring a compact yet powerful model for diverse information extraction applications.
• Use cases where precise and efficient extraction of structured or unstructured information is critical.
• Research and development in natural language understanding focusing on IE tasks.