DrugDiscoveryBench / README.md
zkoch1's picture
Build DrugDiscoveryBench (2026-08-24 12:39 UTC)
e495631 verified
|
Raw
History Blame Contribute Delete
2.01 kB
metadata
license: cc-by-4.0
language:
  - en
pretty_name: DrugDiscoveryBench
tags:
  - drug-discovery
  - biology
  - chemistry
  - benchmark
  - agent
extra_gated_prompt: >
  This dataset contains evaluation-only ground truth and rubrics. Do not
  redistribute them or use them to train or fine-tune models to game the
  benchmark. By requesting access you agree to use the protected fields only for
  evaluation.
configs:
  - config_name: default
    data_files:
      - split: train
        path: train.parquet

DrugDiscoveryBench

Authorized Edison packaging of ScaleAI/DrugDiscoveryBench for agent evaluation. It contains the benchmark's 82 tasks in one train split and preserves the source's evaluation-only access conditions.

Source revision: 10cbbbb5da6f0fa46a6567c7cae0cbb3baa6c7cc.

Capability counts

capability tasks
cheminformatics 10
database_screening 14
molecular_biology 7
patent_mining 13
sar_affinity 7
structural_reasoning 19
target_id_genetics 12

Schema

Each row contains a stable uuid, source question_id and source_category, the agent-visible question, grader_type, protected ground_truth and source metadata in params, and the normalized rubric. input_files is a compact JSON list of task-local filenames. data_storage_uris is a list of matching DEV DSS data_entry: URIs in the same order. The rubric's criteria have stable IDs, correctness or methodology category, and signed numeric point weights.

Outcome criteria use IDs outcome_NNN_positive or outcome_NNN_negative; process criteria use the equivalent process_NNN_* IDs. The suffix is the criterion's signed-point polarity: positive weights award points when met, while negative weights deduct points when the detrimental condition is met.

Load

from datasets import load_dataset

tasks = load_dataset("EdisonScientific/DrugDiscoveryBench", split="train")