yxslpts's picture
Upload 14458 files
b8806f7 verified
|
Raw
History Blame Contribute Delete
2.53 kB

PharmDrugBench Supplementary Appendices

PharmDrugBench task definitions align with four primary papers:

Leaderboard scores are built from appendix tables in AIChemist-Lab repos under sources/:

  • Rx-Bench — primary task metrics from manuscript Tables 2-3 (rx_llm_tables_2_3.csv). The dashboard score is the macro mean across the six primary CMM task metrics, and the leaderboard also splits out all six task-level rows.
  • MedMatch — entity/route accuracy, drug order generation (entity_accuracy_table.csv, route_accuracy_table.csv). evaluation_results.json is included for reference but is not currently consumed by the build script.
  • LLM-DDI — DDI identification Table 3 experiment accuracies (ddi_identification_table3.csv). MedGemma-27B is listed as a separate model row when source tables report MedGemma rather than base Gemma 3 27B.
  • LLM-Uncertainty-DDI — DDI verification accuracy (table_4_results.csv)
  • Pokemon-Drugs-Names — fictitious-drug confabulation rates (PMC Table 2; embedded in build script)

Supported models (7)

GPT-4o-mini, GPT-5 Chat, MedGemma-27B, Gemma 3 27B, Llama 3.3 70B, Qwen3 32B, DrugGPT

Regenerate benchmark data

npm run build:benchmark   # writes server/data/benchmark.json
npm run db:reseed         # reload PostgreSQL from JSON

Run locally

docker compose up -d db    # Postgres on localhost:5447
npm run dev                # API + UI on :8447 with hot reload

Production:

docker compose up -d --build   # http://localhost:8447

Optional: upload DOCX supplements here for manual table extraction:

  • benchmarking-paper-supplement.docx
  • ddi-paper-supplement.docx
  • pokemon-supplement-4.23.docx