thanatchon36/cpilens-reference-set

cpilens: reference set and Thai CPI taxonomy

The evaluation labels and taxonomy tables of the paper How much does product classification cost in basis points?. The paper classifies Thai product text into the items of the Thai consumer price index (base 2566 / 2023), then measures what the classifier's errors do to measured inflation.

📄 Paper · 💻 Code · 🤗 Model · 🔖 DOI 10.5281/zenodo.23686886

The official specification text is not in this dataset. The reference set holds ids, item codes and the sampling design. The Thai government source pages state no licence, so the project does not redistribute their files. See Getting the text.

No training corpus is published here. make data in the code repository rebuilds it from the public sources, with checksums.

Contents

Each CSV file is copied unchanged from commit 1186432 of the code repository, at the same path. The configs load parquet/<config>.parquet, which holds the same rows with the column types listed under Fields, so codes such as 00000 keep their leading zeros.

Config	File	Rows	Content	Licence
reference_set	data/reference_set.csv	3,910	The held-out evaluation rows of the paper: ids, labels and the sampling design.	Attribution, non-commercial research use (no upstream licence stated)
gold_sample	data/gold/sample.csv	600	Thai-market Open Food Facts products drawn for hand labelling. It has no labels and was not used in this version.	ODbL 1.0
item_scope	taxonomy/thcpi_scope.csv	29	The classifier's label set: the in-scope CPI items and their COICOP 2018 class.	MIT
category_crosswalk	taxonomy/off_categories_to_thcpi.csv	119	Crosswalk from Open Food Facts category tags to CPI items, which gives the weak labels.	ODbL 1.0
aggregate_weights	taxonomy/thcpi_weights_2566.csv	29	Official aggregate weights recovered from TPSO's monthly releases, base 2566 (2023).	MIT for the derivation; the official figures are TPSO's (no upstream licence stated)
item_weights	taxonomy/thcpi_leaf_weights_2566.csv	111	Item weights. Item weights inside an official cell are estimates, not official figures.	MIT for the derivation; the official figures are TPSO's (no upstream licence stated)
aggregate_codes	taxonomy/thcpi_weight_aggregates.csv	29	The aggregates in TPSO's weight table, in the order of the table.	MIT for the derivation; the official figures are TPSO's (no upstream licence stated)
source_manifest	data/manifest.csv	28	URL and SHA-256 of every raw source file, which make data checks on download.	MIT

Use

from datasets import load_dataset

ref = load_dataset("thanatchon36/cpilens-reference-set", split="test")  # the reference set, the default config
scope = load_dataset("thanatchon36/cpilens-reference-set", "item_scope", split="train")

The reference set

• Source: the price-collection specifications the Ministry of Commerce publishes as open data, each under the official item code TPSO uses.
• Rows: 2,410 held-out specifications of 26 in-scope items, all kept, and 1,500 held-out out-of-scope specifications in a fixed hash-ordered sample.
• Weighting: weight each row by 1 / incl_prob, so estimates refer to the whole held-out list.
• Split: whole sub-items are held out inside every item. A fifth of an item's sub-items, at least one, are test.
• Coverage: 2 of the 28 in-scope items have no specifications in the published file and cannot be evaluated. Items with few specifications have wide intervals.
• Not sealed: cascade thresholds and classifier settings were chosen on the training and validation splits only, but the development grid logged reference scores for every run. They were visible during development.

For intervals, resample whole cluster values within each stratum. The errors of a method cluster within sub-items.

Getting the text

Scoring a saved prediction needs only this dataset. To run a classifier on the specifications, build them from the source:

git clone https://github.com/thanatchon36/cpilens && cd cpilens
uv sync && make data

data/interim/moc_specs.parquet then holds the text. A row's id here is moc: plus that file's spec_id. If an upstream file has changed since this release, the build stops at the checksum check.

Limitations

• Specifications are the statistical office's own short descriptions, not shop or web listings.
• Open Food Facts is crowd-sourced, and its Thai products served as a development set only.
• There are no prices. The data support classification, not index compilation.
• This is not an official classification. Do not use it as, or to infer, any statistical office's or central bank's internal data or methods.

Licences and attribution

The licence is set per file, as in the table above. The paper's data sources give every source's URL, retrieval date, licence and checksum.

• Open Food Facts and Open Beauty Facts: © Open Food Facts contributors, https://world.openfoodfacts.org, under the Open Database License 1.0. A database derived from these files must carry the same licence.
• Thai government sources: Ministry of Commerce open data, and Trade Policy and Strategy Office (TPSO) publications. Their pages state no licence, so this project treats them under attribution and non-commercial research use.
• COICOP 2018: © United Nations. Class codes and titles are cited.

Fields

reference_set

Column	Type	Meaning
id	string	moc: plus the official 16-digit specification code
label	string	TPSO item code of the in-scope item, or OOS
incl_prob	float64	Inclusion probability; weight a row by its inverse
stratum	string	Bootstrap stratum: the label
cluster	string	Bootstrap cluster: the official 7-digit sub-item
unsure	bool	Always false here; kept for the hand-labelling format

gold_sample

Column	Type	Meaning
order	int32	Labelling order
barcode	string	Product barcode
source	string	off (Open Food Facts) or obf (Open Beauty Facts)
stratum	string	CPI weight cell of the weak label, none or OOS
incl_prob	float64	Inclusion probability within the stratum
double	bool	Selected for a blind second labelling pass
name_th	string	Product name in Thai
name_en	string	Product name in English, if any
brands	string	Brands
quantity	string	Quantity as entered
image_url	string	Link to the front image (CC BY-SA; not copied here)

item_scope

Column	Type	Meaning
label	string	TPSO item code
name_th	string	Item name in Thai
name_en	string	Item name in English
in_scope	bool	Whether the item is in scope
scope_group	string	Group of the item within the scope
coicop2018	string	COICOP 2018 class
coicop_pages	string	Pages of the UN classification that give the class
note	string	Note on the mapping, if any
includes	string	Short description of what the item covers

category_crosswalk

Column	Type	Meaning
tag	string	Open Food Facts category tag
label	string	TPSO item code, or OOS
priority	int32	Priority when a product has several matching tags
basis	string	Reason for the mapping

aggregate_weights

Column	Type	Meaning
code	string	TPSO aggregate code
name_en	string	Aggregate name in English
kind	string	Level of the aggregate
w0	float64	Base weight, the mean across the releases
w0_spread	float64	Largest minus smallest w0 across the releases
w_pr	float64	Price-updated weight share, as used in the paper
release_months_be	string	Releases used, in Buddhist Era years
price_ref_be	string	Price reference month, in Buddhist Era years

item_weights

Column	Type	Meaning
leaf	string	TPSO item code
name_th	string	Item name in Thai
cell	string	Official node whose weight the item shares
residual_cell	bool	Whether official nodes are nested below that node
cell_weight	float64	Weight of the cell
w_equal	float64	Estimated item weight: equal split of the cell
w_nnls	float64	Estimated item weight: non-negative least squares

aggregate_codes

Column	Type	Meaning
row	int32	Row in the published table
code	string	TPSO aggregate code
name_en	string	Aggregate name in English
kind	string	Level of the aggregate

source_manifest

Column	Type	Meaning
path	string	Path under data/raw/
url	string	Source URL
sha256	string	SHA-256 of the file

Citation

@misc{chongmankhong2026cpilens,
  author = {Chongmankhong, Thanatchon},
  title  = {How much does product classification cost in basis points?
            Multilingual classification of Thai product text for consumer price measurement},
  year   = {2026},
  doi    = {10.5281/zenodo.23686886},
  url    = {https://thanatchon36.github.io/cpilens/paper/}
}

The views expressed are the author's own and do not necessarily reflect those of the Bank of Thailand. This work uses only publicly available data. This is independent work, not a BIS, ECB or Bank of Thailand publication.