Laboratory-language grounding on LabGuard-Bench chemistry-held-out hydrochloric acid 1.0
91.8Tag F1LabGuard-Hybrid
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| LabGuard-Hybrid2026.06 | 91.8 | 74.3 | 57.2 | |
| Base Grounder + Constraint decodingBackbone=Qwen3-8B, Decoding Strategy=Constraint decoding2026.06 | 90.2 | 71.5 | 54.8 | |
| Base Grounder + LoRA fine-tuningBackbone=Qwen3-8B, Fine-tuning=LoRA2026.06 | 88.7 | 68.4 | 52.3 | |
| Qwen2.5-72BEvaluation Protocol=0-shot2026.06 | 73.6 | 40.3 | 38.7 | |
| Base GrounderBackbone=Qwen3-8B2026.06 | 68.4 | 32.8 | 28.3 | |
| Regex BL2026.06 | 52.8 | 35.2 | 18.5 |