Curating an ORD dataset

You are preparing a reaction dataset for the Open Reaction Database (ORD) and recording the result on Tokens for Progress (TfP). ORD's own documentation is the authority; this Skill is the route through it.

1. Claim the work where ORD sees it

1. Read the Problem's target input: the paper, its DOI, where the source data is, and the ORD wishlist issue. Read that issue and its comments first; the ORD Program Manager (@bdeadman) records progress and caveats there.
2. On the wishlist board add yourself as Assignee on the card and move it to Reserved (GitHub sign-in needed). Only reserve what you will finish; ORD expects a publishable dataset within about 60 days.
3. If you want the 2026 Dataset Reward ($300 first dataset, $299 second), send the application form named on the rewards page and wait for pre-approval. Eligibility (GitHub account, chemistry background, US sanctions check, at most two rewards per person) is ORD's decision, not TfP's.

2. Build the dataset

• Get the source data machine-readable: SI spreadsheets are best; PDF tables need extraction to CSV; plate heat-map images are not an acceptable source.
• Every compound needs a SMILES or InChI. See ORD's Compound Identifiers guide.
• Use the ORD Reaction Editor to write one reaction template per experimental method and enumerate it over your CSV (Reaction Templates guide). Python (ord_schema) is needed for kinetics (many outcomes per reaction), for hundreds of compounds only named in the source, or for translating from another schema (for SURF files use github.com/bdeadman/surf).
• Record what the paper reports, including negative outcomes, and nothing the paper does not support. Fill provenance (DOI, record creator) for every reaction.
• Validate locally with ord_schema.validations.validate_datasets (pip install ord-schema). Fix every error; read the warnings.
• Share the in-progress dataset with the Program Manager through the editor's groups feature or the issue when unsure; ORD encourages early checks.

3. Submit to ORD

Follow the submission workflow: a pull request to open-reaction-database/ord-data containing the .pb.gz dataset. ORD peer review will ask for corrections; the dataset is accepted when the pull request is merged and the dataset id (ord_dataset-<32 hex>) is live.

4. Submit on TfP

Upload one artifact in role document: the same dataset as protobuf JSON (google.protobuf.json_format.MessageToJson(dataset); ord-schema's message_helpers.write_message(dataset, "x.json") does this). The sandbox stage checks it against document_schema: an object with reactions (at least one), each reaction with inputs, outcomes and provenance, field names either in protobuf JSON camelCase or the proto snake_case, enum values as ORD spells them (SMILES, REACTANT, MILLIMOLE). Unknown fields are rejected. This is a structural check only; ord_schema validation is stricter and the reviewer runs it.

In the Submission summary give: the ord-data pull request URL, the dataset id once assigned, the ORD wishlist issue, and a two-line account of what the dataset covers and what the source did not allow you to capture. Put working notes (extraction scripts, compound identifier tables, open questions) in Attempt Notes and files so the next curator does not repeat them.

Submit on TfP once ORD has accepted the dataset. A Submission made earlier will pass the sandbox stage and fail review on ord_status; it is still a record, and you can submit again after acceptance.

What the reviewer checks

Fidelity to the source (a sample of reactions against the paper and SI), that the JSON passes ord_schema validation, and that the dataset is accepted into ord-data. Declare any agent that assisted you.