# CompleteRXN Benchmark: Completing Unbalanced USPTO Chemical Reactions

## Facts

- Provider: Intelligent Molecular Systems group (CompleteRXN authors)
- License: MIT (https://huggingface.co/datasets/Intelligent-molecular-systems/completeRXN-benchmark/raw/main/README.md)
- Size: 37 files
- Format: CSV, JSON (in sampled files)
- Stored at: Hugging Face, https://huggingface.co/datasets/Intelligent-molecular-systems/completeRXN-benchmark
- Access: public, readable without an account
- Category: Chemistry & Materials
- Source: https://arxiv.org/abs/2605.00222
- Page: https://scigantic.com/archive/93d81088-1344-522e-b152-79d3d1b967c2
- Keywords: Chemistry & Materials, benchmark, CSV, JSON, Hugging Face

## About

206,423 chemical reactions for reaction completion: given a USPTO patent reaction that is missing molecules, so its atoms do not balance, predict the missing species so the completed reaction balances for atoms and charge. Reference completions come from FlowER mechanistic reactions matched to their USPTO records, with stereochemistry restored from USPTO where possible. Reactions are reaction SMILES (reactants>reagents>products) with an incomplete input, a harmonized default input and the balanced target. Three split types test generalization to increasingly unfamiliar reactions: random, group out-of-distribution and extreme out-of-distribution, each with 5 repetitions; the paper evaluates on a filtered test set restricted to tokens in the pretrained Molecular Transformer vocabulary. Listing the repository (measured 2026-11-13) shows benchmark_final.csv (699 MB, all reactions keyed by Unique_identifier), random_splits/ (5 CSV files), group_splits/, extreme_ood_splits/, eligibility/, a croissant.json and raw/ with master_dataset_mapped.csv (1.9 GB) and the FlowER single-reaction TSV. Open with pandas and RDKit, or the completerxn Python package that downloads and evaluates splits. Released under the MIT licence; cite CompleteRXN: Toward Completing Open Chemical Reaction Databases (arXiv:2605.00222).

## Load the data

```python
import pandas as pd
df = pd.read_csv("hf://datasets/Intelligent-molecular-systems/completeRXN-benchmark/benchmark_final.csv")
df.head()
```

## Data structure

File types in sampled files: CSV, JSON

Example file: `benchmark_final.csv`

## Open it

- In a hosted notebook with the data mounted: https://scigantic.com/archive/93d81088-1344-522e-b152-79d3d1b967c2
- Full record as JSON (schema card included): https://api.scigantic.com/api/archives/93d81088-1344-522e-b152-79d3d1b967c2
- Over MCP: `get_schema_card` with id `93d81088-1344-522e-b152-79d3d1b967c2` at https://api.scigantic.com/mcp

## Related topics

- [JSON datasets](https://scigantic.com/topics/json)
- [CSV datasets](https://scigantic.com/topics/csv)
- [Benchmark datasets](https://scigantic.com/topics/benchmark)
- [Chemistry & Materials datasets](https://scigantic.com/topics/chemistry-and-materials)
