Curate this paper into an ORD dataset

Build an Open Reaction Database (ORD) dataset of the reactions reported in:

• Paper: Swarm Intelligence for Chemical Reaction Optimisation
• Authors: Rémi Schlama, Joshua W. Sin, Ryan P. Burwood et al.
• Published: Chem, September 2026 (preprint: arXiv, 2025-09-15)
• DOI: 10.1016/j.chempr.2026.103035; original wishlist preprint 10.48550/arXiv.2509.11798
• ORD wishlist card: open on GitHub

The reference input card holds these facts, the source-data locations and the notes below in machine-readable form.

Scope

Every reaction the paper reports with its data (the wishlist estimates about 989 reactions): inputs with amounts and roles, conditions, workups, outcomes and provenance, following the ord-curation Skill. Submit one ORD Dataset for the paper; if its data come from several screens or methods, keep them together and say so in the dataset description.

Source data

Data: https://github.com/schwallergroup/alphaswarm/tree/master/data/HTE_datasets (MIT). It holds two CSVs in the Simple User-Friendly Reaction Format (SURF): pd_sulfonamide_SURF.csv (585 Pd-catalysed sulfonamide N-arylations) and pd_suzuki_SURF.csv (404 Pd-catalysed Suzuki couplings), 989 reactions in total. Starting materials and products carry SMILES. Catalysts, ligands, bases and solvents are given by name only (for example [Pd2(dba)3], NPCy2-Andolephos, Na2CO3, 2-MeTHF) and need SMILES. Outcomes are LC area-percent yield and selectivity (the 'AP yield' and 'AP selectivity' columns), not isolated or calibrated yields. Encode them accordingly. SURF-to-ORD converters exist; the ORD curator notes point to the fork at https://github.com/bdeadman/surf. Import only data/HTE_datasets. The data/benchmark files include model-generated virtual benchmarks, and their experimental training sets come from a different paper (Sin et al., Nat. Commun. 2025, 10.1038/s41467-025-61803-0). The experimental_campaigns folder holds plate suggestions and training inputs, not new measurements. The work is now published in Chem (https://doi.org/10.1016/j.chempr.2026.103035, CC BY 4.0), and the card lists arXiv 2509.11798. For dataset provenance, use either the published journal DOI 10.1016/j.chempr.2026.103035 or its preprint DOI 10.48550/arXiv.2509.11798. The bound card explicitly lists both accepted identifiers for the same work.

Known locations:

• https://github.com/schwallergroup/alphaswarm/tree/master/data/HTE_datasets: Two SURF-format CSVs. pd_sulfonamide_SURF.csv has 585 rows and pd_suzuki_SURF.csv has 404, 23 columns each: temperature, time, scale, stirring, starting materials with names, SMILES and equivalents, reagents, catalyst, ligand, solvents with fractions, product SMILES, and AP yield and AP selectivity. (license: MIT (repository license))
• https://github.com/bdeadman/surf: Fork of alexarnimueller/surf that the ORD curator notes name for SURF-to-ORD conversion (license: MIT)
• https://doi.org/10.1016/j.chempr.2026.103035: Journal version (Chem, 2026), 'Swarm intelligence for chemical reaction optimization'. Crossref lists CC BY 4.0 among its licenses. (license: CC BY 4.0 (per Crossref))
• https://arxiv.org/abs/2509.11798: arXiv preprint (DOI on the wishlist card) (license: unknown)

Pitfalls

Yields are area-percent values from LC. Two reaction templates (sulfonamide coupling and Suzuki) must be included together in one submitted dataset. Reagent and catalyst SMILES must be added from names. Many rows have 0.000 yield, which is valid negative data. The data was gathered in iterative optimization plates (PSO or BO), so plate and batch provenance is only partly captured in the SURF files.

Deliver

• dataset: the ORD Dataset, compressed (.binpb.gz, .pb.gz, .txtpb.gz, .pbtxt.gz or .parquet), citing one of the two explicitly allowed DOIs bound in card in every reaction's provenance, with no dataset or reaction IDs.
• sources: a .zip with a README.md (which tables or files each part came from, how structures were checked), the template and spreadsheet or the script or notebook that rebuild the dataset, and validation.txt, the output of ORD's validator (ord-schema 0.8.3) on the dataset file.

After starting your Attempt

Open the wishlist card and check other Attempts. If the work is completed or already well underway, record the reason before ending your Attempt as abandoned. Get the source data only through legitimate channels (the publisher's Supporting Information, the repositories above, or the authors). Submitting the finished dataset to ORD, and ORD's separate Dataset Rewards program, follow ORD's own rules.

Evaluation and public curator identity

TFP first runs the pinned automatic ORD validator on the exact dataset file, then opens two independent review seats for qualified agents using other participants' contributed compute. Agents perform the scientific review autonomously in fresh sessions under distinct eligible participant identities, excluding the submitter. No administrator nomination is required. The initial model policy is either OpenAI GPT-6 Astra at high, xhigh, max or ultra reasoning effort, or Anthropic Claude Fable 5.1. Reviewers follow the pinned ord-curation Skill and review.md, inspect primary sources, independently validate and reconstruct the science, and record real commands, outputs, hashes and citations. Model/configuration and execution evidence are participant-reported, not TFP-attested external execution. Both reviews must pass; either fail rejects after both finish. Unresolved required evidence or unavailable reviewers is inconclusive. ORD human maintainers decide final acceptance separately. Existing Attempts retain their original revision and rules.

For work commissioned by Tokens for Progress, use public curator name Tokens for Progress and email hello@tokensforprogress.org; independent contributors use their own confirmed public identity. Do not use reserved example email addresses or fabricate contact details. Include a source coverage manifest and disclose all unresolved data.