---
license: cc-by-4.0
tags:
- chemistry
- molecule-optimization
configs:
- config_name: ind
  data_files:
  - split: train
    path: ind/train.parquet
  - split: test
    path: ind/test.parquet
- config_name: ood
  data_files:
  - split: train
    path: ood/train.parquet
  - split: test
    path: ood/test.parquet
- config_name: all
  data_files:
  - split: train
    path:
    - ind/train.parquet
    - ood/train.parquet
  - split: test
    path:
    - ind/test.parquet
    - ood/test.parquet
---

# MuMOInstruct, hub-native

`NingLab/MuMOInstruct` (`f3ca492caaca6ecb33b1fcff65d882f10acc3297`), restricted to the paper's own IND/OOD evaluation split (the ten declared property combinations) and restructured into `ind`/`ood` subsets, with the unified columns `id`, `query`, `ground_truth`, `response`, `subtask` and `scoring_context` added beside every original one. `ground_truth` is the train split's own `target_smiles` and `response` its tagged form; both are blank for every test row, which upstream carries no target for -- the oracle scores the one candidate's property change against the source molecule, not a stored answer. See `scripts/datasets/README.md` in the AutoDataSci repository for what those columns mean.

Row counts: {"train": 19841, "test": 7810}.

```json
{
  "ind": {
    "train": 14804,
    "test": 5000
  },
  "ood": {
    "train": 5037,
    "test": 2810
  }
}
```