gregb/product-taxonomy-bench

Dataset Summary

product-taxonomy-bench is an anonymised benchmark dataset for predicting Shopify Product Taxonomy categories from Shopify product tags.

This dataset does not include raw product titles, raw tags, or product URLs. Tags are anonymised as tagNNNNNN.

Start Here

• Read this dataset card for the snapshot layout and field definitions.
• Open the benchmark notebook at notebooks/product_taxonomy_bench.ipynb. It defaults to the fixed paper snapshot at revision paper-submission-2026-09-06, with the paper's width-2,000 network and 10,000-iteration ceiling. On the notebook page, use the Hub's Open in Colab button to run it interactively.
• Use the three snapshot folders according to your goal: first1000/ for a tiny sanity-check slice, paper/ for the point-in-time paper snapshot, and latest/ for the rolling benchmark.

Configurations

Three configurations are provided:

Paper snapshot

• paper-2026-02-11T1915Z (created 2026-02-26T17:35:34.843938+11:00; 6,693 products, 2,542 tags, 363 taxonomies; as_of 2026-02-12T06:15:00+11:00)

Latest snapshot

• latest-2026-08-31T1700Z (created 2026-09-01T03:00:03.046187+10:00; 3,367 products, 1,394 tags, 243 taxonomies)

First 1000 snapshot

• first1000-2026-08-31T1700Z (created 2026-09-01T03:00:08.979026+10:00; 1,000 products, 9,260 tags, 309 taxonomies)

Data Fields

Each record corresponds to one product:

• product_id_hash: SHA-256 hash of a canonicalised product URL
• taxonomy_id: Shopify taxonomy GID
• taxonomy_path: Numeric hierarchy path (dot-separated) when available
• taxonomy_name: Human-readable hierarchy name
• cv_fold: 0–4 fold assignment (or null if missing)
• tag_features: list of {tag_id, in_title, title_part, title_position}

Tag semantics are not included; tag_id values are stable only within a snapshot.

Generation

Products were collected by fetching public Shopify product .json endpoints, then joined to the taxonomy label used by the cantbuymelove site. Tags are uppercased and substring-nested tags are filtered before anonymisation. Title overlap positions are computed by case-insensitive substring search and splitting titles on " - " to match the paper’s tag-battle logic. The paper snapshot is generated with a fixed as_of cutoff timestamp.

Frozen paper replication

The paper-submission-2026-09-06 revision contains a self-contained submission package: corrected runtime, pinned dependencies, archived folds, numerical runners, paper figure inputs and a checksum manifest.

Exact numerical replication requires Python 3.11.11 and the frozen dependencies. The notebook checks the interpreter and core scientific versions; Colab's default environment may differ. For exact replication, follow the package's uv venv --python 3.11.11 instructions. Changing the interpreter can change floating-point tie-breaking. The runner sets 12 numerical threads and checks the reference OpenBLAS configuration; changing these settings can also change neural optimisation. Other platforms can explicitly opt into an exploratory run.

The main matrix has 6,693 products, 2,542 tags and 363 classes. The separately released bounded digitwise diagnostic matrix has 6,527 products, 2,747 tags and 308 classes. These populations must not be interchanged. The package documents the stochastic and digitwise adaptations and their limits; the 90% digitwise threshold belongs to this experiment, not to Mihara's paper.

Use terms

This dataset remains marked license: other. Publication of the anonymised research release does not grant a new licence to the underlying catalogue or change existing rights. For reuse terms not expressly provided here, contact Greg Baker through this dataset repository. No product titles, raw tag strings, source URLs or merchant identifiers are included in the new diagnostic matrix.

Citation

Cite the dataset DOI and the frozen revision used for an experiment:

@misc{baker2026producttaxonomybench,
  title={Product Taxonomy Bench: Frozen Submission Release, 6 September 2026},
  author={Baker, Gregory},
  year={2026},
  publisher={Hugging Face},
  doi={10.57967/hf/8413},
  url={https://huggingface.co/datasets/gregb/product-taxonomy-bench/tree/paper-submission-2026-09-06}
}