Distress Detection on Human-annotated French social media tweets 1.0 (Evaluation sample)
79.5AccuracyDistress Reasoner
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| Distress ReasonerParameters=600M, Fine-tuning=Synth. tweets, FR reasoning2026.04 | 79.5 | 82.2 | 73.7 | 77.7 | |
| CamemBERTv2 BaseParameters=110M, Fine-tuning=Synth. tweets2026.04 | 79.3 | 81.6 | 74.3 | 77.7 | |
| ModernBERT LargeParameters=395M, Fine-tuning=Synth. tweets2026.04 | 79.1 | 81 | 74.4 | 77.6 | |
| Mistral 3 LargeParameters=675B, Fine-tuning=—2026.04 | 78.5 | 88.6 | 64.1 | 74.4 | |
| Distress ReasonerParameters=600M, Fine-tuning=Synth. tweets, EN reasoning2026.04 | 78 | 80.7 | 71.9 | 76 | |
| Claude Sonnet 4.5Parameters=—, Fine-tuning=—2026.04 | 72.2 | 75.8 | 62.8 | 68.7 | |
| Gemini 2.5 ProParameters=—, Fine-tuning=—2026.04 | 71.5 | 71.1 | 69.5 | 70.3 | |
| Qwen3.5Parameters=397B, Fine-tuning=—2026.04 | 67.5 | 91.3 | 36.8 | 52.5 |