References and resources

Presentations, papers, code and related work.

Author

Affiliation

Meilame Tayebjee

Innovation Lab, Insee

Presentations

All the team’s presentations are collected at inseefrlab.github.io/codif-ape-prez. The most relevant to this note are listed below.

Table 1: Selected presentations.

Date	Title	Venue
2026-08	Code2Text: generating synthetic descriptions to train automatic coding (M. Morin, in French)	ENSAE internship, Innovation Lab
2026-07	Interpretable text classification: benchmarking lightweight architectures (J. Bernardi)	Internship, Innovation Lab
2026-06	torchTextClassifiers from an MLOps perspective	ISI World Statistics Congress, Malta
2026-06	When Classifications Change but Deadlines Don’t: Scaling Recoding with LLMs	ISI World Statistics Congress, Malta
2026-05	From Accuracy to Trust: practical methods for uncertainty-aware automatic coding	IAOS Conference, Vilnius
2026-03	Classification of business activities by ML: the case of France	Webinars with sub-Saharan Africa and the OECD
2025-12	torchTextClassifiers: a unified framework for text classification in PyTorch	PyData Global
2025-11	LLM-generated training data for a classification change	Journées de méthodologie statistique
2025-11	torchTextClassifiers	Journées de méthodologie statistique
2025-10	torchFastText / torchTextClassifiers	PyData Paris
2025-05	WP10 country project presentation	AIML4OS
2024-05	Retraining strategies for an economic activity classification model	Q2024, Estoril

Papers

• Tayebjee M., Couralet C., Randriamanana N., Pramil J. (2025). fastText, tout juste en production et déjà obsolète ? Présentation du package torchTextClassifiers… Journées de méthodologie statistique.
• Randriamanana N., Faria T. (2025). Changement de nomenclature statistique à l’ère du Machine Learning : réentraînement basé sur les LLM pour la révision de la NAF. Journées de méthodologie statistique.
• Avouac R., Faria T., Comte F. (2025). L’apport des technologies cloud pour industrialiser le processus d’innovation statistique. Insee, Document de travail M2025-05. Section 4 covers APE coding.
• Faria T., Randriamanana N., Seimandi T. (2024). Retraining strategies for an economic activity classification model. Q2024, Estoril.
• AIML4OS WP10 (2025). From text to code (WP10.4): MLOps tutorial. Insee, Statistics Austria, Destatis.

Code

Table 2: Source code.

Repository	Description
codif-ape	Umbrella workspace (submodules)
codif-ape-nace-revision	LLM recoding of the training data from NAF rev. 2 to NAF 2025 (context- vs retrieval-augmented strategies, majority vote)
codif-ape-preprocess	Data cleaning, rule engine, splits
codif-ape-train	Training, evaluation, MLflow packaging
codif-ape-api	FastAPI serving
codif-ape-cd	Kubernetes manifests, Argo CD
torchTextClassifiers	PyTorch text classification library (docs)
GRAAL	LLM agents for classifications (docs)
codif-ape-docs	This website

Bibliography