CREED

Comprehensive Reactant Exhaustive Enumeration Dataset

A large-scale, quality-controlled resource of chemical reactions for single-step retrosynthesis.

---

It pairs target product structures with many plausible reactant sets (one-step disconnections), reflecting that synthesis planning is inherently multi-solution rather than single–ground-truth.

---

Releases

We ship two parallel releases under the CREED name. Both use the same logical layout (one row per product, many reactant candidates). They differ only in whether a ChemCensor-style score is attached to each candidate reaction.

Release	Description
CREED	Full candidate lists without a per-reaction ChemCensor score in the table. Use this when you only need product–reactants structure or will apply your own filtering.
CREED-CCV	CREED–ChemCensor verified: each candidate includes a ChemCensor score (pipeline chemcensor==0.5.2).

The two tables share the same column and struct field names; only chemcensor_score is non-null in CREED-CCV (where the scorer produced a value).

---

Record layout

Each example is one product and an ordered list of reaction candidates (single-step disconnections).

Field	Meaning
product_smiles	Target product as SMILES (non-isomeric).
reactant_candidates	Ordered list of one-step disconnections. Each element is a struct with three fields:

Per-candidate struct (inside reactant_candidates):

Sub-field	Meaning
product_isomeric_smiles	Product isomeric SMILES for that candidate record.
reactants_isomeric_smiles	List of isomeric SMILES for reactants and reagents in that disconnection.
chemcensor_score	In CREED, always null. In CREED-CCV, numeric ChemCensor score when available.

Data are distributed as Parquet shards (part-00000.parquet, part-00001.parquet, …): one logical table per release, sharded across files for I/O and streaming. CREED does not ship Hugging Face–style named splits (no train / validation / test); there is only this single table per release. If you need held-out sets, define them yourself (e.g. random or scaffold-based) on top of the combined rows.

Loading with Hugging Face datasets

Load CREED

from datasets import load_dataset

ds = load_dataset("insilicomedicine/CREED", data_dir="CREED")

Load CREED-CCV (wildcard under that directory)

from datasets import load_dataset

ds = load_dataset("insilicomedicine/CREED", data_dir="CREED-CCV")

Note: Some Hugging Face internals may log the word “split” while building the table; that is not part of the CREED release layout.

---

Dataset statistics

Totals for the full processed tables (all Parquet shards combined). Products = number of rows; reaction candidates = total count of candidate entries across all products.

CREED (no ChemCensor score per candidate)

Metric	Value
Products	1,493,715
Reaction candidates	22,712,526
Mean candidates per product	~15.21
Min / max candidates per product	1 / 152
Parquet shards	8

CREED-CCV (ChemCensor score per candidate)

Metric	Value
Products	698,765
Reaction candidates	6,368,986
Mean candidates per product	~9.11
Min / max candidates per product	1 / 2,364
Parquet shards	4

Note: Top-level rows are keyed by product without isomer-resolved identity at the row level (product_smiles). Isomerically distinct disconnections (different product_isomeric_smiles / reactants_isomeric_smiles tuples) still land in the same row’s reactant_candidates list, so candidate counts can be extremely large for some non-isomeric products.

Citation

@misc{zagribelnyy2026singleanswerenoughrethinking,
      title={When Single Answer Is Not Enough: Rethinking Single-Step Retrosynthesis Benchmarks for LLMs}, 
      author={Bogdan Zagribelnyy and Ivan Ilin and Maksim Kuznetsov and Nikita Bondarev and Roman Schutski
                and Thomas MacDougall and Rim Shayakhmetov and Zulfat Miftakhutdinov
                and Mikolaj Mizera and Vladimir Aladinskiy and Alex Aliper and Alex Zhavoronkov},
      year={2026},
      eprint={2602.03554},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2602.03554}, 
}