anggars/neural-mathrock

Neural Math Rock Multimodal Emotion Dataset

Dataset Description

This corpus is a large-scale multimodal emotion classification dataset specifically developed for Music Information Retrieval (MIR) and emotional computational analysis within complex musical genres, predominantly Math Rock and Midwest Emo. The dataset consists of exactly 4,000 distinct full-length tracks structured directly from the validated metadata registry.

The primary objective of this corpus is to facilitate robust research in automated audio tagging, computational musicology, and cross-modal emotion prediction utilizing state-of-the-art architectures such as MERT for acoustic feature extraction and text-based transformers for metadata alignment.

---

Dataset Structure

Serialization and Ingestion Guidelines

The serialization pipeline is built upon a high-density, memory-insulated sharding architecture consisting of exactly 400 sequential Parquet shards, with each shard containing a fixed block size of 10 tracks.

To ensure execution stability and eliminate Out-Of-Memory (OOM) anomalies during high-throughput ingestion under CPU boundaries, a deterministic byte-level stream constraint is applied. Audio tracks are systematically truncated to a maximum duration of 180 seconds (3 minutes) directly at the low-level binary stream interface prior to PyArrow serialization.

For downstream modeling pipelines, utilizing the lazy evaluation streaming protocol is highly recommended to bypass massive localized memory allocations:

from datasets import load_dataset

# Optimal initialization for large-scale binarized streaming
dataset = load_dataset("anggars/neural-mathrock", split="train", streaming=True)