MGPHot: Expert Music Dataset

• MGPHot is a dataset comprising 58 continuous perceptual attributes with expert annotations capturing detailed musical properties.
• It supports both an autotagging benchmark—using continuous and discretized labels—and a multimodal recommendation framework enriched with audio and lyric embeddings.
• The dataset employs rigorous evaluation protocols and stratified splits to facilitate research in music representation, cross-modal fusion, and improved cold-start recommendations.

MGPHot is an expert-annotated music dataset centered on 58 continuous perceptual attributes spanning seven musicological categories, and it has also become the basis for a multimodal session-based recommendation benchmark in which LastFM-1K is enriched with MGPHot semantic tags, audio embeddings, lyric embeddings, and listening-completion signals (Ramoneda et al., 8 Sep 2025, Kandagatla et al., 28 May 2026). In the autotagging setting, the benchmark is formally defined as $D=\{(x_i,y_i)\}_{i=1}^N$, where $x_i$ is an audio track retrieved via YouTube and $y_i\in\mathbb{R}^{58}$ are expert annotations in $[0,1]^{58}$; after discretization, the same annotations yield binary tags $t_i\in\{0,1\}^{174}$ (Ramoneda et al., 8 Sep 2025). In the recommendation setting, the name “MGPHot dataset” is used for the LastFM-1K session data augmented with MGPHot metadata and additional multimodal signals, with the stated purpose of moving beyond opaque track IDs toward recommendation models that can reason over acoustic, semantic, and engagement information (Kandagatla et al., 28 May 2026).

1. Origin, scope, and conceptual role

MGPHot was introduced as a benchmark with expert musicological annotations, explicitly contrasting such annotations with generic tag datasets based on folksonomies, amateur uploads, or games with a purpose (Ramoneda et al., 8 Sep 2025). The expert descriptors are continuous rather than binary and are intended to capture fine-grained musical properties such as “Harmonic sophistication” and “Aural intensity,” rather than only broad labels such as genre, mood, or instrumentation (Ramoneda et al., 8 Sep 2025).

The later recommendation paper extends this annotation resource into a larger multimodal pipeline by attaching MGPHot semantic tags to tracks in LastFM-1K and combining them with audio embeddings, lyric embeddings, and listening completion ratios (Kandagatla et al., 28 May 2026). The stated motivation is threefold: to improve cold-start recommendations, to ground LLM reasoning in actual song content, and to support diagnostic studies of cross-modal fusion strategies (Kandagatla et al., 28 May 2026).

A useful distinction therefore exists between two related but non-identical objects. The first is the original expert-annotated MGPHot benchmark for music autotagging and representation evaluation (Ramoneda et al., 8 Sep 2025). The second is the “MGPHot-enriched LastFM-1K benchmark,” a session-based recommendation dataset that uses MGPHot as one semantic layer among several (Kandagatla et al., 28 May 2026). This suggests that MGPHot functions both as a standalone annotation benchmark and as a reusable semantic substrate for downstream recommendation research.

2. Annotation schema and semantic representation

The MGPHot schema defines 58 continuous perceptual attributes across seven categories: rhythm, compositional focus, harmony, instrumentation, sonority, vocals, and lyrics (Ramoneda et al., 8 Sep 2025). In the recommendation benchmark description, the same schema is described as 58 “perceptual” tags organized under lyrics, harmony, rhythm, instrumentation, sonority, composition, and vocals, with each attribute rated on a 0–5 Likert scale and then rescaled to $[0,1]$ (Kandagatla et al., 28 May 2026).

The following summary reproduces the field structure as described for the MGPHot-enriched recommendation benchmark.

Category	Example attributes	Type / values
Lyrics	Angry, Sad, Happy, Humorous, Love	float, $[0,1]$
Vocals	Register (low→high), Timbre (thin→full), Breathiness, Smoothness	float, $[0,1]$
Harmony	Tonality (Minor→Major), Harmonic Sophistication	float, $[0,1]$
Rhythm	Tempo Intensity, Swing Feel, Syncopation, Danceability	float, $[0,1]$
Instrumentation	Drum Set, Electric Guitar, Acoustic Guitar, Piano, etc.	float, $x_i$0
Sonority	Live Recording, Acoustic, Electric, etc.	float, $x_i$1
Composition	Focus on Lead Vocal, Focus on Melody, Focus on Lyrics	float, $x_i$2

For the autotagging benchmark, the continuous representation $x_i$3 is referred to as MGPHot-reg, while a discretized representation, MGPHot-tag, expands the 58 indices into 174 binary tags by splitting each continuous index into “Low” $x_i$4, “Moderate” $x_i$5, and “High” $x_i$6 (Ramoneda et al., 8 Sep 2025). The benchmark reports an overall tag density of 58 active tags per track and a level distribution of 12.0% Low, 55.5% Moderate, and 32.5% High (Ramoneda et al., 8 Sep 2025).

The emphasis on continuous expert judgments is central. Unlike coarse label sets, the schema encodes graded intensities over multiple orthogonal musical dimensions (Ramoneda et al., 8 Sep 2025). A plausible implication is that MGPHot is designed not merely for label prediction, but also for probing whether learned music representations preserve musically interpretable structure.

3. Data acquisition, scale, and formats

In the autotagging benchmark, MGPHot contains $x_i$7 tracks (Ramoneda et al., 8 Sep 2025). Audio is not redistributed directly; instead, tracks are retrieved on demand from YouTube using provided URLs and scripts, with users typically downloading 16 kHz or 44.1 kHz audio and converting it to mono (Ramoneda et al., 8 Sep 2025). No raw WAV files are redistributed; the release instead supplies MD5 checksums and a reconstruction script to guarantee a canonical archive (Ramoneda et al., 8 Sep 2025).

The YouTube matching pipeline is described as follows: query YouTube with “artist + title” and keep the top 5 results; obtain a direct hit via RegEx title matching for 72.91%; apply two QWEN2.5_32B verification passes, one using titles and artists only for an additional 22.86% and one using titles, artists, and descriptions for an additional 3.47%; then manually verify the remaining 163 tracks (Ramoneda et al., 8 Sep 2025). Of the resulting matches, 56.43% come from official artist or label channels (Ramoneda et al., 8 Sep 2025).

Licensing differs across components. The original MGPHot annotations are CC BY 4.0 but forbid redistribution of audio derivatives, whereas the autotagging benchmark releases metadata, YouTube URL lists, reconstruction scripts, and precomputed embeddings under CC BY-NC 4.0; users obtain audio under YouTube’s Terms of Service (Ramoneda et al., 8 Sep 2025). By contrast, the MGPHot-enriched LastFM-1K benchmark is released as “MGPHot v1.0” under CC-BY 4.0, with a DOI at https://doi.org/10.5281/zenodo.20431748, and an interactive API is planned via RapidAPI hub in Q4 2026 (Kandagatla et al., 28 May 2026).

For the recommendation-oriented release, the file structure is explicitly specified:

$t_i\in\{0,1\}^{174}$2

This division between annotation metadata, recovered audio references, and downstream embeddings is an important structural feature of the overall MGPHot ecosystem.

4. Multimodal enrichment in the LastFM-1K recommendation benchmark

The recommendation benchmark starts from LastFM-1K and augments it with MGPHot tags, audio embeddings, lyric embeddings, and listening-completion signals (Kandagatla et al., 28 May 2026). The base interaction data are reported as 814 users after filtering out those with fewer than 1,000 plays, 4.21 million listening events (“scrobbles”), 295,957 unique tracks reduced to the top-50,029 most-played catalog, 421,396 sessions produced with a 20-minute inactivity cutoff and with sessions under 10 events discarded, a time span from February 2005 to June 2009, and sparsity of 0.0175 (Kandagatla et al., 28 May 2026).

For the top-50k catalog, the benchmark reports an average track duration of 4.14 minutes, 81 unique genres with Rock at 26%, Electronic at 7%, and Pop at 3%, language coverage of 72.7% English and 27.1% non-English across 42 language codes, and average lyric length of approximately 210 words with $x_i$8 (Kandagatla et al., 28 May 2026).

Two dense embedding streams are extracted per track. Audio embeddings are produced using EnCodecMAE, CLAP, MERT, Music2Vec, and a 63-dimensional handcrafted MFCC+spectral descriptor baseline, with model-dependent output dimension $x_i$9 and formulation

$y_i\in\mathbb{R}^{58}$0

Lyric embeddings are produced using MiniLM, BGE-M3, MPNet, MultiLG, and BERT, with MPNet selected, output dimension $y_i\in\mathbb{R}^{58}$1, and formulation

$y_i\in\mathbb{R}^{58}$2

All of these details are stated explicitly in the recommendation paper (Kandagatla et al., 28 May 2026).

Semantic metadata are generated by querying Azure OpenAI GPT-5 via batch API with a few-shot prompt using two anchor examples, requesting integer 0–5 ratings for all 58 MGPHot attributes and returning only a JSON object; outputs are clipped to $y_i\in\mathbb{R}^{58}$3, then divided by 5 to yield real values in $y_i\in\mathbb{R}^{58}$4 (Kandagatla et al., 28 May 2026). Calibration anchors are described as spanning extremes across instrumentation, vocal character, energy, and production style, in order to ensure use of the full 0–5 range (Kandagatla et al., 28 May 2026). Validation against published MGPHot ground truth yields mean Spearman $y_i\in\mathbb{R}^{58}$5–$y_i\in\mathbb{R}^{58}$6 per category and $y_i\in\mathbb{R}^{58}$7 (Kandagatla et al., 28 May 2026).

Listening completion is encoded per user $y_i\in\mathbb{R}^{58}$8 and track $y_i\in\mathbb{R}^{58}$9 as

$[0,1]^{58}$0

where $[0,1]^{58}$1 is seconds listened and $[0,1]^{58}$2 is track duration (Kandagatla et al., 28 May 2026). Over the top-50k catalog and 4.2 million plays, the mean is approximately 0.78, the median approximately 0.85, and the 10th–90th percentile range is [0.25–0.97], with many skips below 0.2 and most full listens above 0.9 (Kandagatla et al., 28 May 2026).

5. Standardized evaluation protocols

For the autotagging benchmark, standardized train, validation, and test splits are denoted $[0,1]^{58}$3, with approximate cardinalities $[0,1]^{58}$4 (Ramoneda et al., 8 Sep 2025). The splits are constructed by iterative multilabel stratification enforcing four conditions: matching marginal distributions for all 58 descriptors within 2% of global, a similar fraction of official uploads in each split, year-of-release balance, and disjoint main artists across splits (Ramoneda et al., 8 Sep 2025).

The released autotagging benchmark also includes one mean-pooled feature vector per track from six listed state-of-the-art encoders: Whisper, CLAP, MAEST, MERT, MusicFM, and OMAR-RQ, although the text states “seven” and notes that the published repository currently lists six (Ramoneda et al., 8 Sep 2025). Researchers can load these embeddings via a Python utility (Ramoneda et al., 8 Sep 2025).

Two downstream tasks are defined. MGPHot-tag, MagnaTagATune, and MTG-Jamendo are evaluated as multilabel classification tasks with sigmoid outputs and binary cross-entropy loss, whereas MGPHot-reg is evaluated as 58-way regression in $[0,1]^{58}$5 with mean squared error and no sigmoid (Ramoneda et al., 8 Sep 2025). The probe consists of a frozen encoder followed by a two-layer MLP with 512 hidden units and ReLU, trained after mean-pooling over time to obtain one $[0,1]^{58}$6-dimensional vector per track; the optimizer is AdamW with learning rate $[0,1]^{58}$7, weight decay $[0,1]^{58}$8, batch size 128, and early stopping with patience 50 (Ramoneda et al., 8 Sep 2025).

Evaluation emphasizes mean average precision for classification and RMSE for regression. For each label or tag $[0,1]^{58}$9,

$t_i\in\{0,1\}^{174}$0

and

$t_i\in\{0,1\}^{174}$1

Precision@K and recall@K are also computed for analysis but mAP is emphasized (Ramoneda et al., 8 Sep 2025).

In the recommendation benchmark, usage guidance is framed around session-based frameworks and LLM-based sequential reasoning. For SASRec, BERT4Rec, and GRU4Rec, the prescription is to concatenate ID embeddings with audio and lyric embeddings plus metadata, then apply a linear projection into transformer or GRU input space (Kandagatla et al., 28 May 2026). Weighted-sum, cross-attention, or FiLM fusion are explicitly identified as options for experimentation (Kandagatla et al., 28 May 2026). In E4SRec-style LLM-based recommendation, multimodal embeddings are projected into LLM token space via LoRA or an adapter, with zero-shot prompting using metadata and completion ratio as natural-language instructions and fine-tuning on LLaMa-2/3 or Qwen to attend jointly to audio, lyrics, metadata, and engagement signals (Kandagatla et al., 28 May 2026).

6. Empirical findings, interpretation, and research significance

The autotagging benchmark highlights systematic differences between expert and generic annotations. Generic tags are described as coarse and noisy, while expert tags are uniformly annotated across seven musical dimensions and provide fine-grained continuous intensity judgments (Ramoneda et al., 8 Sep 2025). Empirically, on MagnaTagATune and MTG-Jamendo, MAEST leads in mAP with 0.493 and 0.154 respectively, while on MGPHot-tag, CLAP and MERT achieve top scores at approximately 0.375 and MAEST drops to 0.347 (Ramoneda et al., 8 Sep 2025). On MGPHot-reg, RMSE is lowest for CLAP at 0.165 and MERT at 0.164, and Whisper ranks highly in vocal and lyrics dimensions despite being an ASR model (Ramoneda et al., 8 Sep 2025). The paper concludes that no single encoder dominates across all datasets and categories, and that masked-token prediction models such as MERT and OMAR-RQ offer the most balanced performance (Ramoneda et al., 8 Sep 2025).

The recommendation benchmark reports that integrating content-based features improves over ID-only baselines by up to 95% in Recall and 79% in NDCG (Kandagatla et al., 28 May 2026). At the same time, it states that naive multimodal fusion does not always yield additive improvements, which is presented as evidence of challenges in cross-modal integration (Kandagatla et al., 28 May 2026). This is a significant qualification: the benchmark is not merely intended to show that more modalities help, but to expose when and why fusion strategies fail.

A common misconception would be to treat MGPHot as simply another music tagging corpus. The evidence in these papers supports a broader interpretation. In its original form, MGPHot is a benchmark for evaluating whether audio representations capture expert-defined musical attributes (Ramoneda et al., 8 Sep 2025). In its recommendation-oriented extension, it becomes one component of a multimodal sequential decision problem in which semantic descriptors, acoustic embeddings, lyric embeddings, and engagement signals are jointly modeled (Kandagatla et al., 28 May 2026). Another potential misconception is that expert annotation makes generic tag datasets obsolete. The benchmarking results instead emphasize complementarity: category-level analysis shows that “Genre” in generic datasets and “Lyrical content” in MGPHot differ substantially in difficulty, underscoring that the datasets probe different aspects of music understanding (Ramoneda et al., 8 Sep 2025).

Taken together, these uses position MGPHot as an expert-centered infrastructure for research on music representation learning, autotagging, and multimodal recommendation. The original benchmark provides reproducible splits, retrieval scripts, and precomputed embeddings for systematic probing of music encoders (Ramoneda et al., 8 Sep 2025), while the enriched LastFM-1K release provides a large-scale benchmark for testing how content and engagement interact in next-track prediction for both classical session encoders and emerging LLM-based recommenders (Kandagatla et al., 28 May 2026).