Update model card: add July 2026 benchmark re-run (claude-haiku-4-5, omni-moderation, mistral-moderation)

README.md

@@ -154,7 +154,7 @@ Each example carries `sentence`, `label` (ClassLabel: `clean`/`offensive`), `idx
 
 ### Benchmark vs. commercial moderation APIs
 
-From the [companion paper](https://arxiv.org/abs/2505.04640), on the same OMCD_Typica.ai_Mix test split:
+**Original benchmark** — from the [companion paper](https://arxiv.org/abs/2505.04640) (May 2025), on the OMCD_Typica.ai_Mix test split (n = 630):
 
 | Model | Accuracy | Macro F1 | Toxic F1 | Not-Toxic F1 |
 |---|---:|---:|---:|---:|
@@ -163,7 +163,16 @@ From the [companion paper](https://arxiv.org/abs/2505.04640), on the same OMCD_T
 | Mistral (mistral-moderation-latest) | 0.649 | 0.641 | 0.588 | 0.694 |
 | Anthropic Claude (claude-3-haiku-20240307) | 0.659 | 0.617 | 0.743 | 0.492 |
 
-The culturally adapted model achieves the best balance between catching toxic content and avoiding false positives — the general-purpose APIs miss much culturally nuanced toxicity (indirect insults, sarcasm, cultural idioms), while prompt-driven LLM moderation tends to over-flag benign content.
+**Updated re-run** — July 2026, same gold test set (n = 630, balanced), same inputs to all APIs, using each provider's then-current moderation endpoint (weighted precision / recall / F1):
+
+| Model | Precision | Recall | F1-score |
+|---|---:|---:|---:|
+| **Typica.ai (custom BERT-based model)** | **0.832** | **0.830** | **0.830** |
+| Anthropic Claude (claude-haiku-4-5-20251001) | 0.695 | 0.657 | 0.646 |
+| OpenAI (omni-moderation-latest) | 0.692 | 0.630 | 0.607 |
+| Mistral (mistral-moderation-latest) | 0.633 | 0.592 | 0.571 |
+
+Fourteen months after the original benchmark, the performance gap persists even against newer commercial models: the culturally adapted classifier still leads by ~18+ F1 points. General-purpose APIs continue to miss culturally nuanced toxicity (indirect insults, sarcasm, cultural idioms), while the specialized model maintains the best balance between catching toxic content and avoiding false positives.
 
 ## Limitations & Bias
 
@@ -191,21 +200,6 @@ If you use this model, please cite:
 }
 ```
 
-Please also cite the OMCD dataset the training data builds upon:
-
-```bibtex
-@article{essefar2023omcd,
-  title   = {OMCD: Offensive Moroccan Comments Dataset},
-  author  = {Essefar, Kabil and Ait Baha, Hassan and El Mahdaouy, Abdelkader and El Mekki, Abdellah and Berrada, Ismail},
-  journal = {Language Resources and Evaluation},
-  volume  = {57},
-  number  = {4},
-  pages   = {1745--1765},
-  year    = {2023},
-  doi     = {10.1007/s10579-023-09663-2}
-}
-```
-
 ## Contact
 
 **Hicham Assoudi** — Founder & Applied AI Researcher, Typica.ai · PhD (AI/NLP)