FELM logo

FELM

Free

a meta-benchmark that evaluates how well factuality evaluators assess the outputs of large language models (LLMs).

FreeFree tier
Type
Open Source

About FELM

FELM (Factuality Evaluation of Large Language Models) is a meta-benchmark designed to evaluate how well factuality evaluators assess the outputs of LLMs. It addresses the under-explored area of evaluating the evaluators themselves. The benchmark collects responses generated by ChatGPT across five domains: World Knowledge, Science/Technology, Writing/Recommendation, Reasoning, and Math. Each response is annotated with fine-grained factuality labels, resulting in 847 instances with 4,427 labeled segments. FELM provides a leaderboard comparing various evaluators, including GPT-4, Vicuna-33B, and ChatGPT, using metrics such as F1 Score and Balanced Accuracy. The dataset and code are publicly available, and the paper was presented at NeurIPS 2023 Datasets and Benchmarks Track.

Key Features

Covers five distinct domains: World Knowledge, Science/Technology, Writing/Recommendation, Reasoning, and Math
Fine-grained factuality annotation with 4,427 labeled segments across 847 instances
Provides a leaderboard for comparing factuality evaluators (e.g., GPT-4, Vicuna-33B, ChatGPT)
Includes both positive and negative factuality segments for balanced evaluation
Open source with publicly available dataset and code

Pros & Cons

Pros
  • Fine-grained annotation allows detailed analysis of evaluator performance
  • Multi-domain coverage increases benchmark diversity and applicability
  • Public leaderboard facilitates transparent comparison of evaluators
  • Open source dataset and code enable reproducibility and extension
Cons
  • Responses are generated only by ChatGPT, limiting diversity of LLM outputs evaluated
  • Moderate dataset size (847 instances) may not capture all factuality challenges
  • Focuses solely on English text, not tested on multilingual scenarios

Best For

Evaluating and comparing factuality evaluators for LLM outputsResearch on improving factuality detection in large language modelsBenchmarking progress in LLM reliability and truthfulness assessmentTraining or fine-tuning factuality evaluation models