Google Product Taxonomy Synthetic Dataset (GPSD-47k / Pre-Splitted)

Dataset Description

Dataset Summary

GPSD-47K is a fully synthetic, LLM-generated benchmark dataset for hierarchical text classification. It is built on the structural foundation of the Google Product Taxonomy and designed to support research into the effect of extreme taxonomy depth on error propagation in hierarchical classifiers.

The dataset contains 47,390 product descriptions across 1,354 leaf categories, organized in a strict 5-level single-label tree hierarchy — making it one of the deepest tree-structured classification benchmarks publicly available. Unlike established deep benchmarks such as RCV1 or the NYT corpus, which model category structures as directed acyclic graphs (DAGs) for multi-label tasks, GPSD-47K follows a strict single-label tree structure, making it methodologically comparable to benchmarks such as WOS and DBpedia while operating at significantly greater hierarchy depth.

This version of the dataset provides a pre-defined stratified train/validation/test split (70/10/20), recommended for benchmarking and reproducibility. A full unsplit version of the corpus is also available separately.

Supported Tasks

• Hierarchical text classification (single-label, strict tree structure)
• Flat multi-class text classification at any individual hierarchy level
• Evaluation of error propagation across taxonomy depths
• Level-wise classifier evaluation (L1 through L5)

Languages

English

---

Dataset Structure

Hierarchy Overview

The taxonomy follows a strict 5-level tree structure (L1–L5), preprocessed from the original Google Product Taxonomy to retain only paths with exactly five levels. The node distribution across levels is as follows:

Level	Description	Node Count
L1	Top-level category	14
L2	Sub-category	37
L3	Category group	114
L4	Category	266
L5	Leaf class (classification target)	1,354

Data Fields

• text: A natural-language e-commerce product description of 50–80 words, written in a realistic product listing style and covering attributes such as materials, dimensions, compatibility, performance characteristics, and use context.
• l1, l2, l3, l4: Labels for each of the four parent hierarchy levels, enabling level-wise classification experiments.
• l5: The leaf-level classification label and primary classification target.

Data Splits

This version provides a stratified 70/10/20 split, stratified by L5 leaf label to ensure every leaf class is represented across all three subsets. This guarantees reproducibility and cross-study comparability.

Split	Documents
Train	33,173
Validation	4,739
Test	9,478
Total	47,390

---

Dataset Creation

Source Data and Taxonomy

The structural backbone is derived from the publicly available Google Product Taxonomy. The original taxonomy contains variable-length paths; it was preprocessed to retain only paths of exactly five levels (L=5), yielding 1,354 leaf classes across 14 top-level categories. No original Google data or real product listings are contained in this dataset — the taxonomy serves purely as a structural scaffold.

Generation Model

All textual data was generated using Claude Sonnet 4.6 (Anthropic) in a zero-shot setting.

Prompt Design

The generation followed a carefully engineered prompt blueprint with the following core design principles:

Isolated chat contexts per leaf class. To maximize linguistic diversity and counteract style collapse — the tendency of LLMs to produce monotone, repetitive text patterns over long shared contexts — each of the 1,354 leaf classes was generated in a fully isolated chat context with no shared system prompt across classes.

Label leakage prevention. Words appearing directly in the class label (e.g. "FM" or "Transmitter" for the class FM Transmitters) were defined as hard negative constraints in every prompt. This prevents trivially classifiable surface-form patterns from being embedded in the text, ensuring the classification signal must come from semantic content rather than keyword matching.

Enforced feature variance. To prevent one-sided semantic focus within a class, the LLM was instructed to distribute the descriptive focus of generated texts across five distinct functional or technical attribute dimensions (e.g. material, compatibility, dimensions, performance, format/pack) in approximately equal 20% portions. The specific attribute dimensions were tailored to each individual leaf class.

Fixed sample count of 35 per class. Preliminary experiments showed that requesting larger quantities per prompt led to a sharp increase in semantic and syntactic redundancy within the LLM output. 35 samples per class was identified empirically as the optimal trade-off between generation efficiency and intra-class diversity.

Post-hoc Quality Assurance

A two-pass automated deduplication and quality assurance pipeline was applied after generation:

Pass 1 — Intra-category deduplication: All generated documents were vectorized using Term Frequency-Inverse Document Frequency (TF-IDF). Pairwise cosine similarity was computed for all document pairs, grouped by L1 top-level category. Any text pair exceeding a similarity threshold of p > 0.7 was flagged, and an iterative LLM rewrite process was initiated for the flagged documents. This was repeated until no document pair within any L1 group exceeded the threshold.

Pass 2 — Cross-category deduplication: The full TF-IDF vectorization and pairwise cosine similarity procedure was then repeated on the entire corpus without category grouping. This second pass ensures that no accidental lexical or semantic overlap exists across category boundaries, preventing inadvertent data leakage between classes at any hierarchy level — including between semantically adjacent categories sharing a common parent node.

This two-pass procedure guarantees high lexical and semantic heterogeneity both within and across all categories in the corpus.

Class Balance and Distribution

The dataset is perfectly balanced at the leaf level (L5): every one of the 1,354 leaf classes contains exactly 35 samples. Since inner nodes of the taxonomy have varying numbers of child nodes, this bottom-up balance induces a moderate class imbalance at higher hierarchy levels — top-level categories with more descendant paths naturally accumulate more samples than narrower top-level categories.

This design is methodologically deliberate: it dampens the extreme long-tail effect typical of real-world e-commerce catalogues at the leaf level, allowing experimental results to be attributed primarily to the effects of hierarchy depth and error propagation, rather than being confounded by severe data scarcity in individual leaf classes.

---

Limitations and Bias

• Synthetic origin: All texts are LLM-generated and do not reflect real product listings. While the generation process was designed to produce realistic e-commerce language, the lexical and stylistic distribution may differ from real-world product catalogue text in ways that are difficult to fully characterize.
• Label leakage mitigation is heuristic: The forbidden-word constraints reduce but cannot fully eliminate all surface-form cues related to class labels, particularly for multi-word or morphologically derived label terms.
• Taxonomy scope: Only paths of exactly five levels from the Google Product Taxonomy are represented. Product categories with shallower paths are excluded by design.
• English only: The dataset is monolingual and not representative of multilingual e-commerce contexts.
• Generator model bias: As with all LLM-generated datasets, the texts may reflect stylistic biases, knowledge cutoffs, or distributional tendencies of the generating model (Claude Sonnet 4.6).

---

Citation

If you use this dataset in your research, please cite:

@dataset{rollwage2026gpsd,
    author       = { Max Rollwage },
    title        = { gpsd-47k (Revision 879a38f) },
    year         = 2026,
    url          = { https://huggingface.co/datasets/MRollwage/gpsd-47k },
    doi          = { 10.57967/hf/8834 },
    publisher    = { Hugging Face }
}

---

License

CC BY 4.0 — You are free to share and adapt this dataset for any purpose, provided appropriate credit is given.

---

Additional Information

This dataset was created as part of a Bachelor's thesis investigating hierarchical text classification and error propagation across taxonomy depths. The Google Product Taxonomy is the intellectual property of Google LLC and is used here solely as a structural scaffold. No original Google data is contained in this dataset.

The full unsplit corpus is available as a companion dataset at: MRollwage/gpsd-47k-raw