StructLM-7B: Specialized for Structured Knowledge Grounding
StructLM-7B is a 7 billion parameter model from the StructLM series by TIGER-Lab, specifically fine-tuned for structured knowledge grounding (SKG) tasks. It is built upon the CodeLlama-Instruct-hf base model and trained for three epochs on the comprehensive SKGInstruct Dataset, which combines 19 SKG tasks with the SlimOrca dataset.
Key Capabilities & Features
- Structured Data Interpretation: Excels at processing and reasoning over linearized structured inputs, including:
- Tabular data
- Knowledge triples (e.g., from DART)
- Knowledge graph schemas (e.g., from GrailQA)
- Research-Oriented: Primarily intended for research purposes in SKG, offering a specialized tool for complex data interpretation.
- Performance: Demonstrates strong performance across various SKG benchmarks, including ToTTo (49.4), GrailQA (80.4), CompWebQ (78.3), and Spider (72.4) for held-in tasks, and BIRD (22.3), InfoTabs (55.3), and FinQA (27.3) for held-out tasks.
Intended Use Cases
StructLM-7B is ideal for applications that require an AI assistant to analyze and reason over structured information, particularly when the input can be linearized into a specific format. While highly specialized for SKG, it may not be suitable for general chat or other broad applications due to its focused training.