OrganicChemistryReactions

Training-ready dataset for organic reaction-mechanism modeling, optimized for the oMeBench / oMeS metric. The core unit is a typed, mass/charge-balanced elementary mechanistic step with a valid intermediate SMILES — not an overall transformation.

Loading

from datasets import load_dataset
ds = load_dataset("TejasviTyagi/OrganicChemistryReactions", "sft_mechanisms")
# chat SFT rows: parse the JSON-encoded columns
import json
messages = json.loads(ds["train"][0]["messages"])   # [{role, content}, ...]
reference = json.loads(ds["train"][0]["reference"]) # [[subtype, canonical_smiles, weight], ...]

Configurations

• sft_mechanisms — train: 9378, validation: 505
• sft_reactions — train: 6402, validation: 348
• pretrain — train: 2338

A training-ready dataset for a reaction-mechanism LM, optimized for oMeBench/oMeS. The core unit is a typed, balanced elementary mechanistic step with a valid intermediate SMILES.

Headline

• Clean mechanism records (post validate+decon+dedup): 2338
• Train / Val: 2217 / 121 (stratified by dominant subtype + level; 30 strata)
• Test: the official oMe-Gold (196 rxns), held out — never trained on.
• Ontology: 11 types / 31 subtypes (parsed from the benchmark).

Provenance

• template_expanded: 2338

Sources (clean records)

• oMe-Silver: 2338

Difficulty levels

• easy: 422
• medium: 1680
• hard: 236

License inventory

Source	Verdict
Open Reaction Database (ORD) — ~2M react	USABLE
USPTO (Lowe) — "Chemical reactions from	USABLE
RMechDB (Baldi Research Group, UC Irvine	QUARANTINE
Wikipedia named reactions (Tier C) — enc	USABLE
OpenStax Organic Chemistry (Tier C) — CC	QUARANTINE
PMechDB (Public Database of Elementary P	QUARANTINE
Chem LibreTexts (chem.libretexts.org) —	USABLE_FOR_RESEARCH_ONLY
oMeBench (Silver/Gold/Template) — GitHub	USABLE

No paywalled/copyrighted sources are ingested. NC-ND sources (PMechDB, RMechDB) and the no-AI-clause source (OpenStax) are quarantined; LibreTexts (NC) is research-only and kept separable.

Decontamination

• Removed as contaminated vs oMe-Gold+Template: 7
• Post-check gold InChIKey leaks: 0 (0 required).

Ontology coverage (step subtype counts)

subtype	steps
acid_base_proton_transfer	3278
nucleophilic_addition	2090
leaving_group_elimination	1630
nucleophilic_substitution	692
1,2-shift	263
heterolytic_cleavage	203
sigmatropic_rearrangement	201
electrophilic_addition	189
proton_elimination	178
oxidative_addition	177
electrophilic_substitution	139
single_electron_transfer	127
coordination	121
reductive_elimination	76
cycloaddition	74
electrocyclization	71
radical_initiation	63
radical_coupling	59
radical_substitution	52
group_transfer	39
radical_propagation	36
homolytic_cleavage	29
radical_elimination	27
cycloreversion	27
radical_rearrangement	20
dissociation	11
radical_termination	7
cheletropic_reaction	4

Task views (Phase 8)

Known gaps

• Tier-A elementary-step supply is limited to oMe-Silver + template expansion because PMechDB/RMechDB (the spec's primary sources) are CC-BY-NC-ND and cannot be redistributed as a derived corpus.
• Under-covered subtypes (<5 steps) may score poorly on oMeS; consider targeted expansion or licensed sourcing.
• A small set of curated mechanisms whose step 1 introduces reagents disconnected from the substrate (e.g. Appel/NG-007) are rejected by the continuity gate; recoverable with a reagent-introduction step type.
• Atom mapping (RXNMapper) confidence is modest on exotic charged intermediates; low-confidence maps are flagged, not trusted.

Reproduce

make setup   # venv + pinned deps + clone oMeBench
make all     # phases 1-9
make eval-format-check
make test

Decontamination (held-out test integrity)

• Blacklist: oMe-Gold, oMe-Template
• Records removed as contaminated: 7
• Post-check oMe-Gold InChIKey leaks: 0 (0 required — no test record leaks into training)

Intended use

• SFT / continued-pretraining of chemistry LMs for stepwise mechanism prediction; the reference column also drives GRPO against the verifiable oMeS reward.

Out of scope

• Broad chemical knowledge, retrosynthesis planning at scale, or safety-critical use. Scores are highest in-domain (same named reactions, new substituents) and should not be read as general chemical mastery.

Attribution & licenses

Derived from openly-licensed sources only. See license_inventory.json in this repo for the full per-source verdict. oMeBench data/prompts/scorer: MIT (skylarkie/oMeBench). If a rebuild includes ORD (CC-BY-SA-4.0) or USPTO/Lowe (CC0-1.0), the share-alike terms of CC-BY-SA propagate to the combined dataset.