Conditional Semantic Textual Similarity on Re-annotated C-STS 30% hold-out (val)
76.08Spearman CorrelationPoLi-RL (Qwen3-8B)
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| PoLi-RL (Qwen3-8B)Training Paradigm=RL, Parameters=8B2025.10 | 76.08 | 74.16 | |
| QA-SimCSE_LARGETraining Paradigm=SFT, Parameters=355M2025.10 | 72.9 | 72.3 | |
| Qwen3-8BTraining Paradigm=SFT, Parameters=8B2025.10 | 72.09 | 70.41 | |
| GPT-3.5Training Paradigm=Few-shot2025.10 | 66.1 | 64.8 | |
| Qwen3-8BTraining Paradigm=Few-shot, Parameters=8B2025.10 | 64.42 | 64.5 | |
| GPT-4Training Paradigm=Few-shot2025.10 | 64.4 | 60.2 |