---
pretty_name: LMD Deduplication Supplements
dataset_name: lmd-dedup-supplements
language:
  - en
license: cc-by-4.0
size_categories:
  - 100K<n<1M
tags:
  - music
  - midi
  - mir
  - deduplication
  - embeddings
  - lmd
  - lakh_midi_dataset
dataset_summary: >
  Pre-computed embeddings (CAugBERT, CLaMP-1024) for duplicate detection on the
  Lakh MIDI Dataset (LMD-clean and LMD-full). Each folder contains `embeddings.pt`
  and `refs.txt` aligned by row.
citation: |
    @inproceedings{lmd_dedup_2025,
    author    = {Eunjin Choi and Hyerin Kim and Jiwoo Ryu and Juhan Nam and Dasaem Jeong},
    title     = {On the De-duplication of the Lakh {MIDI} Dataset},
    booktitle = {Proceedings of the International Society for Music Information Retrieval Conference (ISMIR)},
    year      = {2025}
    }
---

# LMD Deduplication Supplements

This repository provides pre-computed embedding files extracted from the **Lakh MIDI Dataset (LMD-clean and LMD-full)** using the CAugBERT and CLaMP-1024 models.  
These embeddings were used in our paper:  
**"On the De-duplication of the Lakh MIDI Dataset" (ISMIR 2025)**  
[[Paper]](https://ismir2025program.ismir.net/poster_188.html) | [[GitHub Code]](https://github.com/jech2/LMD_Deduplication)

---

## Contents
Each folder includes:
- `embeddings.pt`: Torch tensor of embeddings (shape: *N × D*)  
- `refs.txt`: List of MIDI filenames corresponding to each embedding row

---

## Usage
```python
import torch

# Load embeddings
embeddings_dir = './lmd_full_to_lmd_full_0.95__caugbert_embedding.npy/'
emb = torch.load(embeddings_dir + "embeddings.pt")

# Load references (MIDI filenames)
with open(embeddings_dir + "refs.txt") as f:
    refs = [line.strip() for line in f]

print(emb.shape, len(refs))
# Example: torch.Size([168662, 512]) 168662
```
For details on how these embeddings are used in duplicate detection and evaluation, please refer to the
`Evaluate` and `De-duplication` sections of the main repository.

# Note
Due to storage limits, we only provide embeddings from CAugBERT and CLaMP-1024 here.
If you need embeddings from other models (e.g., CLaMP-512, MusicBERT, etc.), please contact:
Eunjin Choi (jech@kaist.ac.kr)