Conflict Event Classification on GTD (2,000-row stratified sample)
79.7AccuracyConfliBERT
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| ConfliBERTType=Fine-tuned, Evaluation Protocol=Fine-tuned, Number of Parameters=110M2026.03 | 79.7 | 83.1 | 70.8 | |
| Confli-mBERTType=Fine-tuned, Evaluation Protocol=Fine-tuned, Number of Parameters=177M2026.03 | 75.8 | 79.7 | 58 | |
| ConflLlama-Q8Type=Fine-tuned, Evaluation Protocol=Fine-tuned, Number of Parameters=8B2026.03 | 72.85 | 75.9 | 61 | |
| Gemini 3 FlashType=Commercial API, Evaluation Protocol=Zero-shot, Number of Parameters=200–400B total (est.)2026.03 | 65.85 | 73 | 60.2 | |
| Claude Haiku 4.5Type=Commercial API, Evaluation Protocol=Zero-shot, Number of Parameters=7–20B (est.)2026.03 | 60.95 | 66.4 | 52.6 | |
| DeepSeek V3.2Type=Commercial API, Evaluation Protocol=Zero-shot2026.03 | 56.45 | 66.5 | 49.2 | |
| Qwen2.5Type=Open-source (local), Evaluation Protocol=Zero-shot, Number of Parameters=14B2026.03 | 51.9 | 55.5 | 37.1 | |
| Llama3.1Type=Open-source (local), Evaluation Protocol=Zero-shot, Number of Parameters=8B2026.03 | 31.25 | 34.8 | 21.4 | |
| Gemma2Type=Open-source (local), Evaluation Protocol=Zero-shot, Number of Parameters=9B2026.03 | 31.1 | 33.4 | 21.5 |