Research map / Large language models / Language, safety and interpretability
MID-FiLD: MIDI Dataset for Fine-Level Dynamics
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces MID-FiLD, a 4,422-sample MIDI dataset of professionally composed fine-level dynamics (CC#1 modulation wheel) with four metadata fields, and a REMI-style transformer baseline for generating CC#1 curves.
Problem
Generating human-like music requires fine-level expressive dynamics within notes, but existing symbolic datasets only provide note-level velocity or lack aligned metadata for such expressions.
Approach
A MIDI dataset where professional composers annotate continuous CC#1 (modulation wheel) curves alongside notes and metadata (instrument, track role, mood, min/max CC#1), plus a token-based encoding that conditions a vanilla transformer on notes and metadata to generate CC#1 sequences.
Key results
- Full-metadata transformer reaches RMSE 21.68 versus 41.53 with no metadata on the CC#1 generation task.
- Difference-of-min for the full-metadata model is 10.99 versus 33.92 without metadata.
- Mood classification from generated CC#1 reaches 0.5510 (4Q) and 0.8058 (V) for the full model, approaching ground truth (0.7245, 0.8495 for 4Q, V).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add instrument metadata (ablation) | MID-FiLD Transformer · MID-FiLD test set | RMSE: 29.4925(±18.45) -> 21.6794(±17.79) | Helped |
| add track-role metadata (ablation) | MID-FiLD Transformer · MID-FiLD test set | RMSE: 31.9308(±18.52) -> 21.6794(±17.79) | Helped |
| add mood metadata (ablation) | MID-FiLD Transformer · MID-FiLD test set | A accuracy: 0.6379 -> 0.6553 | Helped |
| add min/max CC#1 metadata (ablation) | MID-FiLD Transformer · MID-FiLD test set | Difference of min.: 29.9745(±25.90) -> 10.9861(±15.29) | Helped |
| add all metadata (ablation) | MID-FiLD Transformer · MID-FiLD test set | RMSE: 41.5318(±19.34) -> 21.6794(±17.79) | Helped |
Surprising findings
- Track role had the largest influence on fidelity, and removing min/max tokens noticeably worsened RMSE.
- The model without mood had the lowest arousal classification score, even below the model with no metadata.
Limitations
- Mood classification and metadata analysis rely on a simple SVM and grouped labels; the proxy evaluation may not reflect perceptual quality.
- Only the modulation wheel CC#1 is annotated, not other expression controls such as CC#11 or pedals.
- Evaluation is purely objective (RMSE and classifier accuracy) without human listening tests.
Experimental setup
Datasets: MID-FiLD
Benchmarks: MID-FiLD test split
Metrics: RMSE difference of min difference of max mood classification accuracy (4Q, V, A)
Baselines: w/o Instrument w/o Track role w/o Mood w/o Min&max w/o Meta
Builds on: REMI ComMU Transformer
Closest papers by meaning
- MIDI-GPT: A Controllable Generative Model for Computer-Assisted Multitrack Music Composition (AAAI 2025 · similarity 0.88)
- Aria-MIDI: A Dataset of Piano MIDI Files for Symbolic Music Modeling (ICLR 2025 · similarity 0.87)
- MuPT: A Generative Symbolic Music Pretrained Transformer (ICLR 2025 · similarity 0.86)
- MusER: Musical Element-Based Regularization for Generating Symbolic Music with Emotion (AAAI 2024 · similarity 0.86)
- MuseControlLite: Multifunctional Music Generation with Lightweight Conditioners (ICML 2025 · similarity 0.86)
- Text2midi: Generating Symbolic Music from Captions (AAAI 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.