LilyBench - Understanding: leaderboard
LilyBench (2026) tests language models on symbolic music written in LilyPond notation. This board is the macro average in percent over eight understanding tasks adapted from ABC-Eval on raw LilyPond scores from the Mutopia collection (bar counting, metadata, bar ordering, next-bar prediction, captioning, composer and genre recognition), with greedy decoding and no chain of thought; higher is better.
Metric: Macro average (0-1 scaled to %) over the eight Mutopia understanding tasks adapted from ABC-Eval to raw LilyPond text (bar count exact match, metadata QA, bar sequencing penalised Kendall tau, next-bar prediction, metadata prediction, music captioning, composer and genre recognition; 4-way multiple choice where applicable, chance 25); greedy decoding, at most 20 new tokens, no chain of thought; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Phi-4 | 65.1 | #1913 |
| 2 | Qwen 2.5 Coder 14B | 64.2 | #3314 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=lilybench-understanding · How It Works · Data refreshed daily, snapshot 2026-11-13.