---
license: cc-by-4.0
language: [en]
pretty_name: DMPKBench
tags: [dmpk, admet, pharmacokinetics, pk, chemistry, benchmark, qa]
configs:
  - config_name: ADMET-MPO
    data_files:
      - {split: train, path: ADMET-MPO/train.parquet}
      - {split: val, path: ADMET-MPO/val.parquet}
      - {split: test, path: ADMET-MPO/test.parquet}
  - config_name: pk-interpretation
    data_files:
      - {split: train, path: pk-interpretation/train.parquet}
      - {split: val, path: pk-interpretation/val.parquet}
      - {split: test, path: pk-interpretation/test.parquet}
---

# DMPKBench

Edison packaging of the public [DMPKBench](https://github.com/GHDDI-AILab/DMPKBench) subsets for agent evaluation.

Two Hugging Face configs (subsets), each with `train` / `val` / `test` splits at ≈70/15/15 (seeded shuffle, seed=42; remainder assigned to `train`):

| subset | train | val | test | total |
|---|--:|--:|--:|--:|
| `ADMET-MPO` | 456 | 97 | 97 | 650 |
| `pk-interpretation` | 71 | 14 | 14 | 99 |

## Subsets

### `ADMET-MPO`
Multiple-choice ADMET/MPO compound-selection questions from [`ADMET-MPO.json`](https://github.com/GHDDI-AILab/DMPKBench/blob/main/public%20subset/ADMET-MPO.json).
Each row asks which of four SMILES is most consistent with a public experimental assay outcome.
Graded with a single `answer` rubric criterion (correct option / SMILES).
`params` stores JSON `domain`, `answer`, `answer_smiles`, and `source_id`.

### `pk-interpretation`
Open-answer PK curve interpretation questions derived from [`Result_Interpretation.json`](https://github.com/GHDDI-AILab/DMPKBench/blob/main/public%20subset/Result_Interpretation.json).
Options are hidden; agents must answer directly and explain using the provided PK data.
Graded with two rubric criteria (`answer`, `explanation`).

## Schema (one row per question)

| column | notes |
|---|---|
| `uuid` | stable primary key |
| `question_id` | `dmpkbench_admet_mpo_<id>` or `dmpkbench_result_interp_<id>` |
| `provenance` | `DMPKBench` |
| `task_type` | `QA` |
| `grader_type` | `rubric` |
| `question` | full self-contained prompt |
| `input_files` / `data_storage_uris` | unused (empty) |
| `reference_file` / `reference_storage_uris` | unused (empty) |
| `params` | optional JSON (ADMET domain / gold letter / SMILES) |
| `rubric` | LLM-judge criteria |

## Load

```python
from datasets import load_dataset

admet = load_dataset("EdisonScientific/DMPKBench", "ADMET-MPO", split="test")
pk = load_dataset("EdisonScientific/DMPKBench", "pk-interpretation", split="test")
```