TIGER-Lab/StructLM-34B

TEXT GENERATIONConcurrency Cost:2Model Size:34BQuant:FP8Ctx Length:32kPublished:Feb 25, 2024License:mitArchitecture:Transformer0.0K Open Weights Cold

StructLM-34B is a 34 billion parameter large language model developed by TIGER-Lab, fine-tuned for structured knowledge grounding (SKG) tasks. Based on CodeLlama-Instruct-hf, this model excels at analyzing and reasoning over structured information, such as tabular data, knowledge triples, and knowledge graph schemas. With a 32K context length, it is specifically optimized for interpreting linearized structured inputs across 19 distinct SKG tasks.

Loading preview...

StructLM-34B: Specialized for Structured Knowledge Grounding

StructLM-34B is a 34 billion parameter large language model from TIGER-Lab, specifically fine-tuned for structured knowledge grounding (SKG) tasks. Built upon CodeLlama-Instruct-hf, this model is designed to process and reason over various forms of structured information, offering a 32K context window for complex inputs.

Key Capabilities

  • Structured Data Interpretation: Excels at understanding and reasoning with linearized structured inputs, including tabular data, knowledge triples, and knowledge graph schemas.
  • Broad SKG Task Coverage: Trained on the comprehensive SKGInstruct Dataset, which combines 19 distinct SKG tasks with SlimOrca data.
  • Performance: Demonstrates strong performance across various SKG benchmarks, including ToTTo, GrailQA, CompWebQ, MMQA, Feverous, Spider, TabFact, and Dart, as well as held-out tasks like BIRD, InfoTabs, FinQA, and SQA.

Intended Use Cases

  • Research in SKG: Primarily developed for research purposes to advance the understanding and application of structured knowledge grounding.
  • Applications Requiring Structured Data Analysis: Suitable for downstream applications that involve interpreting and reasoning with structured data inputs, where adherence to specific output formats is crucial.

Limitations

While specialized for SKG, StructLM-34B may not perform optimally in general chat or other broad applications due to its focused training. Users should adhere to the specific prompt format provided for optimal performance, which includes a system message emphasizing structured information analysis and reasoning.

Finetunes
1 models