Artificial Analysis 音乐 Arena v1.1：千条提示覆盖17种流派，超3.7万盲测投票

全部来源

• Artificial Analysis 音乐 Arena v1.1：千条提示覆盖17种流派，超3.7万盲测投票

  ArtificialAnlys · 主条目

  作者团队介绍 Music Arenas v1.1 更新：新增 1,000 条提示（每个基准 500 条），覆盖 17 种流派；评测改为仅采用招募评估者小组的盲投票，公开投票不计入评分。团队称更新后榜单基于超 37,000 条盲测偏好投票，每个上榜模型均有超 2,000 票；AA-Music-Vocal v1.1 不提供歌词，以衡量模型的端到端歌曲创作能力。

  查看原始帖子 ↗
  查看原文

  Re What's new in v1.1: ➤ We’re introducing a brand new prompt set to our Music Arenas, consisting of 1,000 prompts (500 per benchmark) curated and balanced across 17 genres, ranging from Classical and Jazz & Blues to EDM, Metal and Reggae. ➤ Each prompt tests song structure, mood, instrumentation or vocal style, with longer, more complex prompts that reflect today's prompting behavior and the emerging capabilities of frontier music models. ➤ Panel votes only: ratings come entirely from blind votes by our recruited evaluator panel. Public Arena votes do not count toward the rating. Our updated leaderboards are launching with more than 37,000 blind human preference votes: more than 16,000 on AA-Music-Vocal v1.1 and more than 21,000 on AA-Music-Instrumental v1.1, with over 2,000 votes for every ranked model. On AA-Music-Vocal v1.1 we do not supply lyrics, so the benchmark measures each model's end-to-end songwriting.

前后事件

相关事件