NAF

1 The NAF 2025 nomenclature

The Nomenclature of French Activities 2025 (NAF 2025) is France’s official classification of economic activities, aligned with NACE Rev. 2.1. Every business unit is assigned an APE code from this nomenclature, which underpins official economic statistics across Europe. It organises all activities into five nested hierarchical levels; a full interactive breakdown is available in the appendix.

The training dataset was built by relabelling INSEE’s previous 2.7-million-observation corpus from NACE 2.0 to NACE 2.1. While 75% of codes map one-to-one and were converted by direct table lookup, the remaining 25% are multivocal (one old code → several candidates) and, due to their overrepresentation in practice, account for ~52% of observations. These required reading the activity description to resolve. A Rule-Based Augmented Generation (RBAG) pipeline was developed for this: given a description and its legacy NACE 2.0 code, the correspondence table retrieves the 2–5 candidate NACE 2.1 codes, their official explanatory notes are injected into a prompt, and a LLM picks the best match. Evaluated on a 30,000-observation benchmark annotated by ~25 experts (strict inter-annotator agreement: 62%), an ensemble of three LLMs reaches 78% accuracy — with much of the gap attributable to genuine ambiguity rather than model errors. The final corpus contains ~2.3 million NAF 2025 labels, though only ~1.3 million have an already-preprocessed text description usable for training — plenty on its own, training on the full 2.3 million wouldn’t have been practical anyway. More details in the codif-ape-nace-revision repository.

NAF 2025 — Hierarchical structure

Click on a level to display its description.

Section
21
codes
20 in training
▼
1 letter  ·  A to U
21 broad thematic families (agriculture, industry, services…). Highest level of the nomenclature, standard granularity for macroeconomic analysis.
≈ 4 divisions per section
▼
Division
87
codes
84 in training
▼
2 digits  ·  01 to 99
First numeric level. Breaks each section into homogeneous activity domains; reference granularity for national and European sectoral statistics.
≈ 3.3 groups per division
▼
Group
287
codes
274 in training
▼
3 digits  ·  XX.X
Refines the division into sets of activities sharing common economic characteristics. Intermediate level for detailed sectoral analysis.
≈ 2.3 classes per group
▼
Class
651
codes
586 in training
▼
4 digits  ·  XX.XX
Detail level shared with the European NACE Rev. 2.1. Defines a specific activity; the level at which statistical harmonisation across EU member states takes place.
≈ 1.1 sub-classes per class
▼
Sub-class
747
codes
679 in training
▼
4 digits + letter  ·  XX.XXY
Finest level, specific to France. May split or merge classes to reflect national specificities. The trailing letter distinguishes sub-classes within the same class.

2 Two modelling approaches

Two independent modelling approaches are covered, each on its own page:

Baseline
Single-level classification
Predicts only the finest-grained sub-class code (679 classes) directly, in a single pass.
Explore single-level analysis →
Multi-task
Multi-level classification
Predicts all five NAF levels at once with a single model: one shared token embedder feeds five independent per-level classification heads, trained jointly on one harmonised loss.
Explore multi-level analysis →

3 Detailed breakdown of the NAF 2025

Select a section to explore the hierarchy.
Section 21
Division —
← Select a section
Group —
← Select a division
Class —
← Select a group
Sub-class —
← Select a class
Groups and classes are identified by their code only (labels not available at these levels).