Beyond Traditional Benchmarks: Analyzing Behaviors of Open LLMs on Data-to-Text Generation

TL;DR

This paper investigates how open large language models perform on data-to-text generation tasks, highlighting their fluency but also significant semantic accuracy issues, and introduces a new data collection tool to avoid benchmark contamination.

Contribution

The study provides a comprehensive analysis of open LLMs' behaviors on data-to-text tasks and introduces Quintd, a novel tool for collecting uncontaminated structured data.

Findings

Open LLMs generate fluent, coherent texts in zero-shot settings.

Over 80% of outputs contain at least one semantic error.

Semantic accuracy remains a major challenge for open LLMs.

Abstract

We analyze the behaviors of open large language models (LLMs) on the task of data-to-text (D2T) generation, i.e., generating coherent and relevant text from structured data. To avoid the issue of LLM training data contamination with standard benchmarks, we design Quintd - a tool for collecting novel structured data records from public APIs. We find that open LLMs (Llama 2, Mistral, and Zephyr) can generate fluent and coherent texts in zero-shot settings from data in common formats collected with Quintd. However, we show that the semantic accuracy of the outputs is a major issue: both according to human annotators and our reference-free metric based on GPT-4, more than 80% of the outputs of open LLMs contain at least one semantic error. We publicly release the code, data, and model outputs.