chiral-carbon / kg-for-science

Knowledge Base for scientific publications, powered by LLMs
MIT License
1 stars 0 forks source link

Improve the current pred performance on the dev set #3

Open chiral-carbon opened 1 month ago

chiral-carbon commented 1 month ago

Currently the best metrics are around: { "precision": 0.4329, "recall": 0.3974, "f1": 0.4144, "union_precision": 0.5864, "union_recall": 0.5216, "union_f1": 0.5521, "avg_time_per_sentence": 4.0315, "total_time": 463.6508 }

We want to improve these values. This could be achieved by improving the prompt quality, the schema or other things.

chiral-carbon commented 1 month ago

try reasoning or adding constituency tests (will slow down inference)