FACTS Grounding V1

We show this table for reference; we do not rank on it.

Data verified Archived version 1

The archived FACTS Grounding V1 protocol tests factual responses grounded in supplied documents. Its original judge setup and results remain separate from Grounding V2.

Evaluation results published by Google DeepMind, Google Research, and Kaggle. How Kaggle Benchmarks reruns tests.

The exact version 1 page declares Apache 2.0. We retain aggregate measurements, attribution, and source links. No tasks, questions, answers, documents, or other underlying dataset contents are included. Original article terms.

Should you trust a FACTS Grounding V1 score?

Read with care: scores cluster at the top.

Saturation Different source configurations: These archived version 1 protocols remain separate from the current version 2 boards. The complete official download includes configurations hidden by default source filters. Source identities are preserved, and these measurements do not enter overall or category rankings.
Who produced the rows Mirrored table: Every row is mirrored from Google DeepMind, Google Research, and Kaggle as it publishes it; BenchLM did not re-run these results. Source.
Freshness and weight Display only: Current, refreshed quarterly. Display only; it does not move rankings. Public benchmark set.

Row statuses come from the verification registry; saturation and freshness from the benchmark lifecycle. A provider self-report is displayable evidence, not an independent run.

Published score (%) on FACTS Grounding V1 — Archived version 1

We mirror the published published score (%) view for FACTS Grounding V1. Llama 3 - Grounded Language Model has the highest reported value at 89.4%. These archived version 1 protocols remain separate from the current version 2 boards. The complete official download includes configurations hidden by default source filters. Source identities are preserved, and these measurements do not enter overall or category rankings. We do not use these results to rank models overall.

Llama 3 - Grounded Language Model

Archived Kaggle benchmark v1; Task revision 1; Evaluated 2025-08-20

89.4%

Gemini 2.5 Pro

Archived Kaggle benchmark v1; Task revision 1; Evaluated 2025-06-06

87.8%

GPT-5

Archived Kaggle benchmark v1; Task revision 1; Evaluated 2025-08-11

87.4%

35 source rowsReasoningCurrentDisplay onlyUpdated Archived version 1

Published score (%) table (35 source rows)

Published score (%) results for FACTS Grounding V1
Rank	Model / configuration	Score	Parameters (B)	Open / closed
1	Llama 3 - Grounded Language ModelCreator not reportedArchived Kaggle benchmark v1; Task revision 1; Evaluated 2025-08-20	89.4%	Not reported	Unknown
2	Gemini 2.5 ProCreator not reportedArchived Kaggle benchmark v1; Task revision 1; Evaluated 2025-06-06	87.8%	Not reported	Unknown
3	GPT-5Creator not reportedArchived Kaggle benchmark v1; Task revision 1; Evaluated 2025-08-11	87.4%	Not reported	Unknown
4	Gemini 2.5 FlashCreator not reportedArchived Kaggle benchmark v1; Task revision 1; Evaluated 2025-07-14	86.4%	Not reported	Unknown
5	Gemini 2.0 FlashCreator not reportedArchived Kaggle benchmark v1; Task revision 1; Evaluated 2025-05-08	85.6%	Not reported	Unknown
6	Claude 3.5 SonnetCreator not reportedArchived Kaggle benchmark v1; Task revision 1; Evaluated 2025-05-08	83.3%	Not reported	Unknown
7	Gemini 2.0 Flash LiteCreator not reportedArchived Kaggle benchmark v1; Task revision 1; Evaluated 2025-06-20	83.2%	Not reported	Unknown
8	DeepSeek-R1Creator not reportedArchived Kaggle benchmark v1; Task revision 1; Evaluated 2025-06-04	82.4%	Not reported	Unknown
9	Gemini 1.5 FlashCreator not reportedArchived Kaggle benchmark v1; Task revision 1; Evaluated 2025-05-16	82.4%	Not reported	Unknown
10	Gemini 1.5 ProCreator not reportedArchived Kaggle benchmark v1; Task revision 1; Evaluated 2025-05-08	81.2%	Not reported	Unknown
11	Gemma 3 27BCreator not reportedArchived Kaggle benchmark v1; Task revision 1; Evaluated 2025-05-08	80.6%	Not reported	Unknown
12	Gemma 3 12BCreator not reportedArchived Kaggle benchmark v1; Task revision 1; Evaluated 2025-05-08	80.5%	Not reported	Unknown
13	Gemini 1.5 Flash 8BCreator not reportedArchived Kaggle benchmark v1; Task revision 1; Evaluated 2025-05-16	79.7%	Not reported	Unknown
14	Claude Sonnet 4Creator not reportedArchived Kaggle benchmark v1; Task revision 1; Evaluated 2025-05-23	79.1%	Not reported	Unknown
15	Claude 3.7 SonnetCreator not reportedArchived Kaggle benchmark v1; Task revision 1; Evaluated 2025-05-08	78.8%	Not reported	Unknown
16	o1Creator not reportedArchived Kaggle benchmark v1; Task revision 1; Evaluated 2025-05-08	78.8%	Not reported	Unknown
17	Claude 3.5 HaikuCreator not reportedArchived Kaggle benchmark v1; Task revision 1; Evaluated 2025-05-08	78.1%	Not reported	Unknown
18	Claude Opus 4Creator not reportedArchived Kaggle benchmark v1; Task revision 1; Evaluated 2025-05-23	77.7%	Not reported	Unknown
19	o3 miniCreator not reportedArchived Kaggle benchmark v1; Task revision 1; Evaluated 2025-05-08	77.6%	Not reported	Unknown
20	Grok 2Creator not reportedArchived Kaggle benchmark v1; Task revision 1; Evaluated 2025-05-08	76.2%	Not reported	Unknown
21	GPT-4oCreator not reportedArchived Kaggle benchmark v1; Task revision 1; Evaluated 2025-05-08	75.9%	Not reported	Unknown
22	Grok 3Creator not reportedArchived Kaggle benchmark v1; Task revision 1; Evaluated 2025-05-29	75.7%	Not reported	Unknown
23	Grok 3 MiniCreator not reportedArchived Kaggle benchmark v1; Task revision 1; Evaluated 2025-05-26	75.0%	Not reported	Unknown
24	GPT-4.1Creator not reportedArchived Kaggle benchmark v1; Task revision 1; Evaluated 2025-05-08	74.5%	Not reported	Unknown
25	Mistral Large 2Creator not reportedArchived Kaggle benchmark v1; Task revision 1; Evaluated 2025-05-08	72.6%	Not reported	Unknown
26	Mixtral 8x22BCreator not reportedArchived Kaggle benchmark v1; Task revision 1; Evaluated 2025-05-08	70.1%	Not reported	Unknown
27	o3Creator not reportedArchived Kaggle benchmark v1; Task revision 1; Evaluated 2025-05-08	69.6%	Not reported	Unknown
28	GPT-4o miniCreator not reportedArchived Kaggle benchmark v1; Task revision 1; Evaluated 2025-05-08	68.4%	Not reported	Unknown
29	Gemma 3 4BCreator not reportedArchived Kaggle benchmark v1; Task revision 1; Evaluated 2025-05-08	64.8%	Not reported	Unknown
30	DeepSeek-V3Creator not reportedArchived Kaggle benchmark v1; Task revision 1; Evaluated 2025-05-08	63.9%	Not reported	Unknown
31	o1 miniCreator not reportedArchived Kaggle benchmark v1; Task revision 1; Evaluated 2025-05-08	62.1%	Not reported	Unknown
32	o4 miniCreator not reportedArchived Kaggle benchmark v1; Task revision 1; Evaluated 2025-05-08	62.1%	Not reported	Unknown
33	CodestralCreator not reportedArchived Kaggle benchmark v1; Task revision 1; Evaluated 2025-05-08	59.7%	Not reported	Unknown
34	Ministral 8BCreator not reportedArchived Kaggle benchmark v1; Task revision 1; Evaluated 2025-05-08	56.1%	Not reported	Unknown
35	Ministral 3BCreator not reportedArchived Kaggle benchmark v1; Task revision 1; Evaluated 2025-05-08	52.2%	Not reported	Unknown

How to read FACTS Grounding V1

The version 1 download supplies overall, public, and private scores. Version 2 changes the evaluation setup, so the archived scores remain separate. This board measures support from the supplied document; it does not establish factual accuracy outside that document.

The download was acquired on the recorded retrieval date. Per-result evaluation timestamps are publisher dates, not evidence of when this leaderboard snapshot was first available.

These archived version 1 protocols remain separate from the current version 2 boards. The complete official download includes configurations hidden by default source filters. Source identities are preserved, and these measurements do not enter overall or category rankings.

Snapshot

35 source rowsArchived v1 protocolComplete downloadDisplay only

Official archived Kaggle leaderboard Official version 1 leaderboard JSON Evaluation method Benchmark-page Apache 2.0 declaration

35 source rows are shown for FACTS Grounding V1. The benchmark falls in the Reasoning category. We keep external benchmark mirrors separate from the weighted global scoring system, so these results remain source-specific evidence. FACTS Grounding V1 is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.

About FACTS Grounding V1

Year

2025

Tasks

1,719 examples: 860 public and 859 private

Format

Document-grounded answering with an ensemble of three judges

FACTS Grounding V1 tests responses to requests accompanied by documents of up to 32,000 tokens. Its 1,719 examples are split into 860 public and 859 private examples. The original protocol uses an ensemble of three model judges and filters responses for eligibility before assessing factual grounding. The version 1 download supplies overall, public, and private scores. Version 2 changes the evaluation setup, so the archived scores remain separate. This board measures support from the supplied document; it does not establish factual accuracy outside that document. These archived version 1 protocols remain separate from the current version 2 boards. The complete official download includes configurations hidden by default source filters. Source identities are preserved, and these measurements do not enter overall or category rankings.

FACTS Grounding V1 methodology Public benchmark source

Freshness and provenance

Version

FACTS Grounding V1 2025

Refresh cadence

Quarterly

Staleness state

Current

Question availability

Public benchmark set

CurrentDisplay only

BenchLM uses freshness metadata to decide whether a benchmark should still be treated as a strong differentiator, a benchmark to watch, or a display-only reference. For the full scoring policy, see the BenchLM methodology page.

Questions

What does FACTS Grounding V1 measure?

The archived FACTS Grounding V1 protocol tests factual responses grounded in supplied documents. Its original judge setup and results remain separate from Grounding V2.

Which source row has the highest reported FACTS Grounding V1 value?

Llama 3 - Grounded Language Model has the highest reported FACTS Grounding V1 value at 89.4%. These archived version 1 protocols remain separate from the current version 2 boards. The complete official download includes configurations hidden by default source filters. Source identities are preserved, and these measurements do not enter overall or category rankings. BenchLM shows this benchmark for display only and does not use it in overall rankings.

How many models are evaluated on FACTS Grounding V1?

The Archived version 1 snapshot contains 35 AI models.

Last updated: Archived version 1 · mirrored from the public benchmark leaderboard

Know when it’s worth switching models

The model to choose, the cheaper alternative, and the release we would wait on.

Read a sample issue

Join 6,100+ subscribers.

One email each week. Unsubscribe anytime.