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

Our complete benchmark and on-line leaderboard supply a much-needed measure of how precisely LLMs floor their responses in offered supply materials and keep away from hallucinations

Giant language fashions (LLMs) are remodeling how we entry data, but their grip on factual accuracy stays imperfect. They will “hallucinate” false data, significantly when given advanced inputs. In flip, this could erode belief in LLMs and restrict their functions in the actual world.

In the present day, we’re introducing FACTS Grounding, a complete benchmark for evaluating the flexibility of LLMs to generate responses that aren’t solely factually correct with respect to given inputs, but additionally sufficiently detailed to supply passable solutions to consumer queries.

We hope our benchmark will spur industry-wide progress on factuality and grounding. To trace progress, we’re additionally launching the FACTS leaderboard on Kaggle. We’ve already examined main LLMs utilizing FACTS Grounding and have populated the preliminary leaderboard with their grounding scores. We are going to preserve and replace the leaderboard as the sphere advances.

FACTS Grounding dataset

To precisely consider the factuality and grounding of any given LLM, the FACTS Grounding dataset contains 1,719 examples, every fastidiously crafted to require long-form responses grounded within the context doc offered. Every instance contains a doc, a system instruction requiring the LLM to solely reference the offered doc, and an accompanying consumer request.

All examples are divided right into a “public” set (860) and a “personal” (859) held out set. We’re releasing the public set as we speak so anybody can use it to judge an LLM. In fact, we all know that problems with benchmark contamination and leaderboard hacking are essential to guard towards, so following commonplace {industry} follow, we’re conserving the personal analysis set held out. The FACTS leaderboard scores are the typical efficiency throughout each private and non-private units.

To make sure a range of inputs, the FACTS Grounding examples embody paperwork with a wide range of lengths, as much as a most of 32,000 tokens (roughly 20,000 phrases), masking domains equivalent to finance, expertise, retail, drugs, and legislation. The consumer requests are equally large ranging, together with requests for summarization, Q&A technology, and rewriting duties. We didn’t embody any examples that would require creativity, arithmetic, or advanced reasoning – capabilities which could require the mannequin to use extra superior reasoning along with grounding.

Collective judgement by main LLMs

To succeed on a given instance, an LLM should synthesize the advanced data within the doc and generate a long-form response that’s each a complete reply to the consumer request and absolutely attributable to that doc.

FACTS Grounding evaluates mannequin responses mechanically utilizing three frontier LLM judges — particularly Gemini 1.5 Professional, GPT-4o, and Claude 3.5 Sonnet. We chosen a mixture of various judges to mitigate any potential bias of a decide giving increased scores to the responses produced by a member of its personal mannequin household. The automated decide fashions had been comprehensively evaluated towards a held-out take a look at set to seek out the perfect performing judging immediate templates and to confirm settlement with human raters.

Every FACTS Grounding instance is judged in two phases. First, responses are evaluated for eligibility, and disqualified in the event that they don’t sufficiently deal with the consumer’s request. Second, responses are judged as factually correct if they’re absolutely grounded in data contained within the offered doc, with no hallucinations.

With the eligibility and grounding accuracy of a given LLM response evaluated individually by a number of AI decide fashions, the outcomes are then aggregated to find out if the LLM has handled the instance efficiently. The ultimate rating for the general grounding job is the typical of all decide fashions’ scores throughout all examples. Discover extra particulars of our FACTS Grounding analysis methodology in our paper.

FACTS Grounding will proceed to evolve

We’re aware that benchmarks might be shortly overtaken by progress, so this launch of our FACTS Grounding benchmark and leaderboard is only the start. Factuality and grounding are among the many key elements that may form the longer term success and usefulness of LLMs and broader AI methods, and we goal to develop and iterate FACTS Grounding as the sphere progresses, regularly elevating the bar.

We encourage the AI neighborhood to engage with FACTS Grounding, consider their fashions on the open set of examples or to submit their fashions for analysis. We imagine that complete benchmarking strategies, coupled with steady analysis and improvement will proceed to enhance AI methods.