AI BriefingLG AI Research: Introducing ChartInstruct Research
Key point
We introduce ChartInstruct, a general-purpose chart vision-language model that applies Instruction Tuning to maximize chart understanding and reasoning capabilities.
Details
At ACL 2024, developing general-purpose models that perform various tasks using chart data has emerged as a major topic. This article focuses on the ChartInstruct model, which analyzes visual information in charts and performs complex reasoning.
To secure data, the researchers utilized web crawling and public datasets, extracting data tables with Gemini Pro Vision and generating questions via GPT-3.5/4. Through this process, they built a total of 191,000 instructions and 70,882 chart data samples.
The model architecture consists of three core components: a Vision Encoder, an Adapter Module, and an LLM. Instead of the conventional CLIP, it uses UniChart, a chart-specific encoder, and performs alignment through an adapter that maps visual features into the language model's embedding space.
Experimental results demonstrate that ChartInstruct achieves superior performance over the existing open-source model UniChart across various benchmarks, including ChartQA, OpenCQA, Chart-to-Text, and ChartFC.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Apple Releases LensVLM-9B for Processing Compressed Text Images