LF-Amazon-131K

An extreme multi-label text classification dataset from the Extreme Classification Repository. Amazon products are labeled with fine-grained category tags drawn from a vocabulary of 131,073 labels.

Dataset Description

Each example is an Amazon product with a title and description, tagged with one or more labels from a massive label space. Labels themselves have rich text metadata (title and description), making this dataset suitable for both traditional extreme classification and zero-/few-shot approaches.

Property	Value
Train examples	294,805
Test examples	134,835
Total labels	131,073
Mean labels per example	~2.3
Median label frequency	3

Label distribution

The label distribution is extremely long-tailed:

Training examples per label	# Labels	% of total
1	25,253	19.3%
2–5	67,439	51.5%
6–10	26,080	19.9%
11–50	11,926	9.1%
51+	375	0.3%

Over 70% of labels have 5 or fewer training examples, making this dataset a challenging benchmark that blurs the line between supervised classification and zero-/few-shot retrieval.

Fields

Field	Type	Description
uid	string	Amazon ASIN
title	string	Product title
content	string	Product description
label_ids	list[int]	Label indices (into the 131K label vocabulary)
label_titles	list[string]	Human-readable label titles
relevance_scores	list[float]	Relevance score per label (all 1.0 in this dataset)
filter_label_ids	list[int]	Curated/filtered label subset indices
filter_label_titles	list[string]	Human-readable filtered label titles

Additional files

• label_vocab.json — ordered list of all 131,073 label titles (index = label ID)
• label_metadata.jsonl — full label metadata (id, uid, title, content) for zero-shot / label-text-aware approaches

Usage

from datasets import load_dataset

ds = load_dataset("Tellurio/LF-Amazon-131K")

example = ds["train"][0]
print(example["title"])
print(example["label_titles"])  # e.g. ["Methodical Bible Study"]

# Label IDs are also available for efficient encoding
print(example["label_ids"])     # e.g. [4315]

Loading label metadata for zero-shot approaches

Each of the 131,073 labels has rich text metadata (title and description) stored in label_metadata.jsonl. This is useful for zero-shot or label-text-aware approaches where you match product text against label text.

import json
from huggingface_hub import hf_hub_download

# Download the label metadata file from the repo
path = hf_hub_download(
    repo_id="Tellurio/LF-Amazon-131K",
    filename="label_metadata.jsonl",
    repo_type="dataset",
)

# Load into a list (index = label ID)
labels = []
with open(path) as f:
    for line in f:
        labels.append(json.loads(line))

# Look up a label by ID
print(labels[4315])
# {"id": 4315, "uid": "...", "title": "How to Read the Bible as Literature", "content": "..."}

Citation

If you use this dataset, please cite the Extreme Classification Repository:

@misc{xmlrepo,
  author = {Manik Varma},
  title = {Extreme Classification Repository},
  url = {http://manikvarma.org/downloads/XC/XMLRepository.html}
}

License

This dataset is redistributed from the Extreme Classification Repository. Please refer to the original source for licensing terms.