npc/isic-activity-coder
ISIC Activity Coder
Accuracy @1 (4-digit)
0.83
Accuracy @3 (4-digit)
0.94
Accuracy @1 (2-digit)
0.92
Coder-agreement rate
0.89
About this model
Classifies free-text business activity descriptions in Arabic or English to ISIC Rev.4 codes at the 4-digit class level, with calibrated confidence and top-3 alternatives. Built to automate coding in establishment surveys and commercial-registration processing.
Intended use
Automated activity coding in statistical production and registry cleaning at NPC and licensing entities; low-confidence cases route to human coders.
Training data lineage
Trained on 310,000 manually coded activity descriptions aligned with establishment-census-aggregates classification audits.
Limitations & bias notes
Multi-activity descriptions (e.g. trading plus contracting) receive only the dominant code, and holding-company boilerplate is a persistent confusion source. Newly emerging digital activities map to legacy ISIC classes imperfectly.
Evaluation metrics
| Accuracy @1 (4-digit) | 0.83 |
| Accuracy @3 (4-digit) | 0.94 |
| Accuracy @1 (2-digit) | 0.92 |
| Coder-agreement rate | 0.89 |
Try it
Live sandboxFeedback
Owning entity
- Updated
- 2026-03-11
- Latest version
- 1.2.0
- License
- Government Internal Reuse
- Access
- Open
Trained on
establishment-census-aggregates →