anggars/neural-mathrock

Neural Math Rock Multimodal Emotion Dataset

Dataset Description

This dataset is a large-scale, fine-grained multimodal emotion classification corpus specifically tailored for music information retrieval (MIR) and emotional computational analysis within complex musical genres, predominantly Math Rock and Midwest Emo. It contains exactly 48,000 localized audio segments derived from an initial master collection of 4,000 distinct full-length tracks.

The primary purpose of this dataset is to facilitate research in computational musicology, automated audio tagging, and cross-modal emotion prediction using state-of-the-art architectures such as MERT-v1 for audio feature extraction and XLM-RoBERTa for metadata text fusion.

---

Dataset Structure

Layout and Block Alignment

The dataset utilizes a highly optimized Temporal Block Sharding Architecture. The distribution matrix is strictly balanced and configured into 12 distinct temporal segments (clips) per track. To optimize sequential data loading and prevent tensor dimension mismatches during batching operations, the records are organized as follows:

• Rows 1 – 4,000: Contain clip_index 1 for all 4,000 tracks, ordered systematically by the master metadata registry.
• Rows 4,001 – 8,000: Contain clip_index 2 for all 4,000 tracks, maintaining the identical master sequence.
• This uniform block-sharding sequence loops continuously up to Block 12 (Rows 44,001 – 48,000), creating a symmetric $4000 \times 12$ matrix topology.

Field Definitions

Every data record within the Parquet shards adheres strictly to the following 7-column schema definitions:

1. artist (string): The verified name of the musical group or performing artist.
2. song (string): The original title of the musical composition.
3. file_name (string): The validated physical filesystem name of the source audio block, completely normalized to prevent unicode character corruption (Mojibake resistance).
4. audio (audio): A structured Hugging Face native audio object containing:
   • bytes: The raw uncompressed biner format of the localized 15-second audio slice.
   • path: The relative path reference of the waveform.
5. emotion (string): The categorical fine-grained emotional label assigned to the specific clip (e.g., neutral, joy, relief, sadness, desire, etc.).
6. clip_index (int64): The sequential slice order tracking number within the temporal matrix bounds (strictly ranging from 1 to 12).
7. confidence (float64): The statistical probability score or algorithm consensus value representing the annotation reliability index.

---

Data Distribution and Characteristics

Class Imbalance Insights

The underlying categorical annotations feature a fine-grained 28-class emotional framework exhibiting a long-tail distribution typical of real-world acoustic data. The majority class is anchored by neutral (15.24%) and joy (10.61%), while nuanced classes like realization, gratitude, and confusion populate the tail end of the spectrum.

Researchers utilizing this corpus are advised to apply loss-weight scaling or specialized cross-entropy optimization penalties to adjust for class density variance:

$${\text{Weight}}_c=\frac{N_{\text{samples}}}{N_{\text{classes}}\times {\text{Count}}_c}\backslash text\{Weight\}\_\{c\}\; =\; \backslash frac\{N\_\{\backslash text\{samples\}\}\}\{N\_\{\backslash text\{classes\}\}\; \backslash times\; \backslash text\{Count\}\_\{c\}\}$$

Structural Consistency

Cross-tabulation audits confirm that the relative density of emotion classes is evenly distributed across all 12 temporal clip_index blocks. This ensures that the sequential split partitions do not introduce demographic or covariate shifts during multi-stage downstream feature fusion training loops.