Haocheng1/mech-infer-train-v1

mech-infer-train v1 — molecularly-diverse mechanism training set

Reaction-mechanism trajectories assembled to broaden the chemical-space and reaction-type coverage of the SFT corpus toward the FukuyamaBench organic-synthesis genre, addressing the weak Set B/C performance.

This is the pre-annotation trajectory dataset. Each record is a reaction with elementary steps (reactants / products SMILES) and conditions, in the schema consumed by annotate_reaction.py → convert_to_train.py → SFT. Run those to produce the Gemini forward-predictive CoT and TRL messages.

Composition (10,000 reactions)

Source	Count	Granularity	Reaction-type labels	Notes
oMe_silver	2,063	multi-step (mean ~4 steps)	yes (reaction_types)	organic-synthesis mechanisms; covers the failing classes
ART_grossman_augmented	956	multi-step	no	functional-group analogs of Grossman textbook mechanisms
ART_grossman_original	17	multi-step	no	original Grossman mechanisms
human_benchmark	274	multi-step	no	Clayden/literature pathways
pmechdb_singlestep	6,690	single elementary step	no	PMechDB polar mechanisms; volume top-up

Multi-step "core" = 3,310 reactions (oMe_silver + ART + HumanBenchmark). The 6,690 PMechDB entries are single-step polar mechanisms (different task granularity, no pericyclic/radical); they are tagged distinctly so they can be re-weighted or excluded (--no-pmechdb). For a pathway-focused model, treat the core as primary.

Failing-class step coverage in the labeled core: pericyclic 389, rearrangement 293, cleavage 215, radical 100, electron-transfer 114.

Provenance & non-contamination

• Independent of FukuyamaBench. Sources are oMe (silver split only; the 196 oMe_gold cases are excluded), Grossman's Art of Writing Reasonable Organic Reaction Mechanisms, HumanBenchmark, and PMechDB. The Fukuyama workbook is never used as a seed.
• Decontamination (exclusion only). Every reaction is dropped if any substantive molecule (≥8 heavy atoms) exactly matches (InChIKey) or is a near-duplicate (Morgan r=2, 2048-bit, Tanimoto > 0.7) of any FukuyamaBench molecule. 1,103 reactions were removed this way (notably real oMe/PMechDB↔ FukuyamaBench substrate overlaps). Fingerprints are used only to exclude near-duplicates — never to select or construct benchmark look-alikes.

Reproduce

python rebuttal/new_figures/generate_art_augmented.py        # ART analogs
python rebuttal/new_figures/build_training_dataset.py --target 10000

Known limitations

• Scaffold diversity of the ART portion is bounded (~11 augmentable textbook scaffolds, mono/light-double substitution only).
• PMechDB single-step entries dominate by count; weight accordingly.
• The pericyclic/radical coverage comes mainly from oMe (484/210 raw steps); adding the cluster-generated pericyclic/named mechanisms would strengthen it.