CommunityFact (Closed-Input) - Finance (Japanese): leaderboard

CommunityFact Finance Japanese asks LLMs to label recent finance claims in Japanese, built from helpful X Community Notes, as True or False from the claim and its date alone, without web search (CommunityFact, 2026). The score is macro-F1 over the two classes on the temporally held-out test claims, in percent, higher is better.

Metric: Macro-F1 (%) over True and False claims on the temporally held-out Finance Japanese test slice (80 True and 47 False standalone claims built from helpful X Community Notes, labels grounded in the note), closed-input (the model sees only the claim and its timestamp, no web search), zero-shot verdict prompt in the claim language; open models greedy, closed models at default settings with reasoning effort medium in their thinking runs; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 10 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1Gemini 2.5 FlashNon-reasoning81.6#463
2aya-expanse-32B79.13#1968
3Grok 4.3Medium78.48#342
4aya-expanse-8B74.97#4141
5Qwen 3 32BNon-reasoning73.03#972
6EuroLLM-22B-Instruct-251266.88#2008
7Qwen 3 14BNon-reasoning64.53#1269
8GPT-5 Nano (2025-08-07)Medium61.8
9EuroLLM-9B-Instruct-251248.22#2674
10Ministral-8B-Instruct-241034.44#4555

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=communityfact-closed-input-finance-japanese · How It Works · Data refreshed daily, snapshot 2026-11-21.