Start with the workflow, not one licence label

A chemistry AI resource may combine source code, trained model files, example data, full datasets and hosted services. A licence label attached to the repository is not necessarily evidence for every component you download or use. Begin by listing the components your planned workflow actually needs, then record a separate source for each.

This guide organises evidence rather than providing a legal opinion. It does not certify scientific performance, compatibility or permission for a particular use. Decisions with material licensing consequences need appropriate review.

Build a component inventory

Describe the intended activity before collecting licence texts: internal evaluation, training, fine-tuning, distributing checkpoints, publishing processed data or operating a service. These activities involve different inputs and outputs, so an inventory should follow the whole workflow.

Use one row per component, with fields for:

  • Identity: resource name, component, release or dataset identifier, and source URL.
  • Intended use: what you will do with it and who will receive the outputs.
  • Evidence: applicable licence file, published terms, notices and provenance records.
  • Status: documented conditions, missing evidence, conflicts and questions requiring review.

The inventory's inputs are the proposed workflow and its source materials. Its output is an evidence record, not a blanket approval. Keep code, checkpoints, data and service access in separate rows even when they share a download location.

Match the evidence to the component

Prefer the official licence file or published terms associated with the relevant release. Repository visibility and downloadable files alone do not establish usage rights. Record the release or revision and the observation date in your working notes; do not substitute a current repository label for evidence about an older artifact.

The resource references illustrate why this distinction matters:

Resource Evidence described in the resource reference Separate question to investigate
MACE Code is listed as MIT; code and checkpoint licences must be considered separately. What terms apply to the exact pretrained checkpoint selected?
GemNet The repository licence is listed as Hippocratic License Version 2.0. Which components does the applicable text cover, and what conditions affect the proposed use?
LocalRetro Repository code is described as CC BY-NC-SA 4.0. What permissions apply to external reaction datasets and model weights?
Therapeutics Data Commons Repository code is listed as MIT. What terms apply to each selected dataset?
Open Reaction Database Helper code is described as Apache-2.0; datasets and descriptive metadata as CC-BY-SA-4.0. What attribution and redistribution conditions apply to the selected data?

These are draft-described distinctions, not findings from a completed licensing review. Confirm the applicable text before making a decision.

Record conditions without filling gaps

Capture licence identifiers and explicit conditions accurately. Keep a short paraphrase for navigation, but retain a link to the controlling text so reviewers can inspect it directly. Do not turn missing information into a claim of unrestricted commercial use.

Useful decision questions include:

  • Does the text explicitly cover this file or dataset?
  • Does it address the planned use or redistribution?
  • Are notices, attribution or other conditions specified?
  • Do repository documentation and artifact-specific terms disagree?
  • Is clarification needed from an authoritative source?

When documents conflict, record both sources and the unresolved issue. A sensible tracking status is “evidence incomplete,” rather than an unsupported legal conclusion.

Follow third-party material and derived outputs

Check whether examples, datasets or pretrained artifacts incorporate material from other sources. Preserve provenance, citations and notices while preparing an internal evaluation. A top-level software licence may not cover bundled or externally retrieved material.

For LocalRetro, the resource reference describes template extraction from reaction training data, preprocessing into CSV files, training checkpoints and decoding into candidate reactants. That sequence suggests several inventory rows; it does not establish permissions for all resulting artifacts.

For Therapeutics Data Commons, dataset loaders and evaluation utilities are distinct from the datasets they retrieve. For Open Reaction Database, keep helper code, reaction records and descriptive metadata distinguishable in your notes. Conversion into another format is a processing step, not evidence that the original conditions disappear.

Worked planning example: a hypothetical reaction workflow

Suppose a research team proposes evaluating LocalRetro and inspecting Open Reaction Database records alongside its reaction-data preparation. This is a hypothetical planning exercise, not a tested integration or a claim that the resources are interchangeable.

  1. Define the boundary. The initial activity is internal evaluation. Sharing data or checkpoints is a separate decision.
  2. List inputs. Record LocalRetro code, the chosen training data, template files, any downloaded checkpoint, and the selected Open Reaction Database files.
  3. Map planned outputs. Include preprocessed CSV files, extracted templates, a trained checkpoint, raw predictions, decoded reactants and the evaluation report.
  4. Attach evidence. Link component-specific licence texts and provenance records. Keep external dataset permissions unresolved until evidence is obtained.
  5. Ask the release question. Before sharing an output, identify the material it contains and route unclear conditions for review.

The practical result is a traceable inventory and a list of open questions. Scientific evaluation and permission decisions remain separate workstreams.

Recheck releases and service terms

Licences and access terms can change between versions. Retain historical records, inspect new releases and review hosted-service terms separately from source code. Access availability is also distinct from permission: a reachable download does not resolve licensing questions.

This process has limits. Resource licence labels may not establish artifact-specific coverage, and documentation may omit terms for weights or third-party data. Update published facts only after new evidence has been checked; leave unknown conditions visible meanwhile.

Sources and a short checklist

Use the relevant repository and documentation as starting points:

Before proceeding:

  • List every required component and planned output.
  • Attach release-specific evidence and preserve notices.
  • Mark gaps and conflicting documents explicitly.
  • Seek appropriate review before consequential use or sharing.
  • Recheck evidence when components or intended activities change.

Continue with resources and scientific tasks. Use source links as evidence starting points, not as certification of permissions or scientific suitability.