Adversarial safety and pedagogical evaluation on SHAPE
-2.72Delta SafeSocraticLM
Evaluation results
| Method | Links | ||
|---|---|---|---|
| SocraticLMScale=8B, Fine-tuned=Tutoring-specific2026.04 | -2.72 | -63.49 | |
| EduchatScale=8B, Fine-tuned=Tutoring-specific2026.04 | -8.84 | -18.29 | |
| GPT-5Variant=Base2026.04 | -16.86 | -7.85 | |
| Claude 4.5Variant=Haiku2026.04 | -31.59 | -38.23 | |
| GPT-5Scale=mini2026.04 | -58.67 | -1.07 | |
| Qwen3Scale=32b2026.04 | -61.76 | -68.91 | |
| GPT-5Scale=nano2026.04 | -62.23 | -0.76 | |
| Claude 4.5Variant=Opus2026.04 | -69.59 | 3.36 | |
| EduchatScale=32B, Fine-tuned=Tutoring-specific2026.04 | -83 | -6.49 | |
| Gemini 2.5Variant=Flash2026.04 | -87.18 | -7.29 | |
| Gemini 2.5Variant=Flash-Lite2026.04 | -87.65 | 3.17 | |
| Qwen3Scale=8b2026.04 | -90.26 | -73.95 | |
| Gemini 2.5Variant=Pro2026.04 | -94.77 | -52.8 | |
| Qwen3Scale=80B2026.04 | -99.29 | -70.33 |