mmarkusmalone/albums_with_moods_contrastive

Dataset Details (Card Organization by HuggingFace)

Dataset Description

This dataset contains synthetic natural-language music queries paired with albums, designed for fine-tuning contrastive and embedding-based retrieval models. Queries are generated from album reviews using curated adjective and music-descriptor vocabularies, enabling semantic alignment between descriptive text and musical works.

The dataset emphasizes mood, texture, and stylistic descriptors rather than only genre labels, making it suitable for music search, recommendation, and creative retrieval tasks.

• —Curated by: Maya Markus-Malone
• —Language(s) (NLP): English
• —License: CC BY-NC 2.0

Dataset Sources [optional]

Built off of: (https://huggingface.co/datasets/mattismegevand/pitchfork)

Uses

This dataset is intended for:

• —Fine-tuning contrastive / dual-encoder models
• —Music semantic search and retrieval
• —Album and artist recommendation systems
• —Query–item embedding alignment
• —Research on text-to-music representation learning

It is particularly well-suited for models trained with:

• —InfoNCE or similar contrastive losses
• —In-batch negatives
• —Dense retrieval architectures

Out-of-Scope Use

This dataset is not suitable for:

• —Audio signal modeling or music generation
• —Supervised classification tasks without modification
• —Commercial applications (due to non-commercial license)
• —Demographic or user-behavior analysis
• —Claims about listener preferences or popularity

Dataset Creation

Curation Rationale

The dataset was created to address the lack of high-quality query–item pairs for training music retrieval models. Human annotation at scale is expensive, and genre labels alone fail to capture the nuanced ways people describe music. By grounding synthetic queries in real album reviews, the dataset balances:

• —Scalability
• —Semantic accuracy
• —Expressive, user-like language

Source Data

The base album reviews were sourced from the Pitchfork dataset:https://huggingface.co/datasets/mattismegevand/pitchfork

Data Collection and Processing

1. 1.Album reviews were ingested from the Pitchfork dataset.
2. 2.Two auxiliary vocabularies (adjectives and music descriptors) were generated using ChatGPT.
3. 3.Reviews were processed to extract relevant descriptor terms.
4. 4.Natural-language queries for each album/review were synthetically generated by combining extracted descriptors.
5. 5.The query/source album pair form contrastive training examples.
6. 6.No scraping or personal data collection was performed.

Who are the source data producers?

The original reviews were written by Pitchfork contributors.

This dataset was built from the following: https://huggingface.co/datasets/mattismegevand/pitchfork

Annotations [optional]

No manual annotation was performed. All queries and descriptors were generated programmatically using large language models and rule-based filtering grounded in review text.

Personal and Sensitive Information

The dataset does not contain personal, sensitive, or private information. All content relates to publicly released music and professionally written album reviews.

Bias, Risks, and Limitations

• —The dataset reflects the editorial biases of Pitchfork reviews.
• —Language and descriptors may overrepresent Western indie and popular music.
• —Synthetic queries may lack the full diversity of real user search behavior.
• —Descriptors are text-based and do not reflect actual audio features.

Recommendations

Users should:

• —Avoid treating descriptors as objective musical truths
• —Be cautious when generalizing beyond Western music contexts
• —Consider augmenting with user-generated queries or audio features
• —Evaluate models on real-world retrieval benchmarks when possible

Dataset Card Authors [optional]

Maya Markus-Malone

Dataset Card Contact

mmarkusmalone@gmail.com