Sign up to save tools and stay up to date with the latest in AI
bg
bg
1

FACTS Grounding: A new benchmark for evaluating the factuality of large language models

Dec 18, 2024 - deepmind.google
The article introduces FACTS Grounding, a benchmark designed to evaluate the factual accuracy and grounding of large language models (LLMs) in their responses to user queries. This benchmark aims to address the issue of LLMs generating false information, or "hallucinations," by assessing their ability to produce factually accurate and detailed responses based on provided source material. The FACTS Grounding dataset includes 1,719 examples requiring long-form responses, divided into public and private sets, and covers various domains such as finance, technology, and medicine. The benchmark uses three LLM judges—Gemini 1.5 Pro, GPT-4o, and Claude 3.5 Sonnet—to evaluate model responses for eligibility and factual accuracy.

The article also announces the launch of the FACTS leaderboard on Kaggle, which tracks the progress of LLMs in grounding their responses. The leaderboard scores are based on the average performance across both public and private datasets. The authors emphasize the importance of continuous improvement in factuality and grounding as key factors for the future success of LLMs and AI systems. They encourage the AI community to engage with FACTS Grounding by evaluating their models on the open set of examples or submitting their models for evaluation. The initiative is led by a team of researchers and acknowledges contributions from various individuals and supporters.

Key takeaways:

```html
  • FACTS Grounding is a new benchmark designed to evaluate the factual accuracy and grounding of large language models (LLMs) in their responses to user queries.
  • The benchmark includes a dataset of 1,719 examples, divided into public and private sets, to ensure comprehensive evaluation and prevent benchmark contamination.
  • Model responses are judged by three frontier LLMs—Gemini 1.5 Pro, GPT-4o, and Claude 3.5 Sonnet—to ensure unbiased evaluation and agreement with human raters.
  • The FACTS Grounding benchmark and leaderboard aim to drive industry-wide progress in improving the factuality and grounding of LLMs, with ongoing updates and community engagement encouraged.
```
View Full Article

Comments (0)

Be the first to comment!