Datasets:

AsciiMAster
/

off-categories-bge-m3

	---
	license: odbl
	language:
	- en
	- de
	- fr
	- pl
	pretty_name: Open Food Facts Categories with bge-m3 Embeddings
	size_categories:
	- 10K<n<100K
	task_categories:
	- feature-extraction
	- sentence-similarity
	tags:
	- open-food-facts
	- food
	- taxonomy
	- embeddings
	- bge-m3
	- multilingual
	- pgvector
	configs:
	- config_name: default
	data_files: off-categories-bge-m3.parquet
	---

	# Open Food Facts Categories with bge-m3 Embeddings

	A flat, embedded version of the [Open Food Facts](https://world.openfoodfacts.org/data)
	category taxonomy. Each row is one OFF category with multilingual display
	names, its full ancestor path, and a 1024-dimensional embedding produced by
	[`BAAI/bge-m3`](https://huggingface.co/BAAI/bge-m3).

	The intent is to let downstream projects do retrieval over OFF categories
	without having to re-fetch `categories.json`, walk the parent graph, or run
	the embedding model.

	## Source project

	Generated by [BetterCategories / offcat](https://github.com/AsciiMAster/BetterCategories),
	a small Go service that does natural-language food category search backed by
	this exact table.

	## Schema

	\| Column \| Type \| Notes \|
	\|-----------------\|---------------------\|--------------------------------------------------------------------\|
	\| `id` \| string \| OFF category id, e.g. `en:apple-juices`. Primary key. \|
	\| `name_en` \| string (nullable) \| English display name. \|
	\| `name_pl` \| string (nullable) \| Polish display name. \|
	\| `name_de` \| string (nullable) \| German display name. \|
	\| `name_fr` \| string (nullable) \| French display name. \|
	\| `parents` \| list<string> \| Direct parent category ids. \|
	\| `ancestor_path` \| string (nullable) \| Pre-computed flattened ancestor chain used as embedding context. \|
	\| `context_hash` \| string (nullable) \| SHA-256 of the text fed to the embedding model. Useful for diffs. \|
	\| `embedding` \| list<float32>[1024] \| `bge-m3` embedding of `name_en + " > " + ancestor_path` (cosine). \|

	Row count: 14352.

	## Embedding details

	- Model: `BAAI/bge-m3` (served via an OpenAI-compatible local API in the
	upstream project; any deployment of the same model produces compatible
	vectors).
	- Dimension: 1024.
	- Distance: cosine similarity is the intended metric.
	- Input text: the English category name concatenated with `" > "` and the
	flattened ancestor path. See `taxonomy.go` in the source project for the
	exact `embedTextFor` formatting.

	## Quick start

	```python
	import pyarrow.parquet as pq
	import numpy as np

	table = pq.read_table("off-categories-bge-m3.parquet")
	df = table.to_pandas()

	mat = np.stack(df["embedding"].to_numpy()) # (14352, 1024)
	mat /= np.linalg.norm(mat, axis=1, keepdims=True)

	# encode a query with bge-m3, normalise, then:
	# scores = mat @ query_vec
	```

	For Postgres users, the same data is what `BetterCategories` writes into a
	`pgvector` `vector(1024)` column with an HNSW cosine-ops index.

	## Licence and attribution

	This dataset is a derivative database of Open Food Facts and is therefore
	distributed under the Open Database License (ODbL) v1.0, the same
	licence as the upstream taxonomy.

	- Upstream data: Open Food Facts, https://world.openfoodfacts.org/data,
	ODbL-1.0.
	- Modifications applied: parent graph flattened into an `ancestor_path`
	string per category; `bge-m3` embeddings computed over the concatenation
	of the English name and that ancestor path; rows with no embedding are
	excluded.

	If you redistribute this dataset or a derivative of it (including dumps,
	indexes, or further-trained models that incorporate the data), you must:

	1. Keep the resulting database under ODbL-1.0.
	2. Credit Open Food Facts contributors.
	3. Note that the data has been modified.

	The Go source code that produced this dataset is licensed under MIT and is
	not covered by ODbL.