victorlxh/iKG-v1.0
The victorlxh/iKG-v1.0 is a 7 billion parameter auto-regressive language model developed by Xiaohui Li, fine-tuned from Vicuna-7B (derived from Meta's LLaMA) with a 4096-token context length. It is specifically designed for knowledge graph construction (KGC) tasks, excelling at extracting structured information in triplet format from text based on specialized instruction-following prompts. This model outperforms GPT-3.5 and Vicuna-7B in KGC, demonstrating capabilities comparable to GPT-4 for this specific application.
Loading preview...
iKG-v1.0: Specialized Knowledge Graph Generation
The victorlxh/iKG-v1.0 is a 7 billion parameter instruction-following language model, developed by Xiaohui Li, specifically fine-tuned for knowledge graph construction (KGC) tasks. Built upon LMSYS's Vicuna-7B, which itself is based on Meta's LLaMA architecture, iKG-v1.0 leverages a 4096-token context length to process and extract structured information.
Key Capabilities
- Instruction-Following KGC: Generates knowledge graphs in a triplet format based on specialized prompts, extracting entities and relationships from input text.
- Customizable Extraction: Supports custom entity categories (e.g., ORG, PERSON, COMP, FIN_INSTRUMENT) and a predefined set of relationship verbs (e.g., Has, Announce, Impact).
- Entity Disambiguation: Capable of consolidating different phrases or acronyms referring to the same entity.
- Performance: Outperforms GPT-3.5 and the base Vicuna-7B model in KGC tasks, showing comparable performance to GPT-4 for this specific application.
Good For
- Researchers and Data Scientists: Ideal for those focused on natural language processing and automated knowledge graph construction.
- Specialized Information Extraction: Particularly effective for extracting structured data from documents where specific entity types and relationships are required.
- Prompt-Engineered KGC: Users can define custom extraction rules via prompt engineering to suit various domain-specific needs.