Dataset Viewer
The dataset could not be loaded because the splits use different data file formats, which is not supported. Read more about the splits configuration. Click for more details.
Couldn't infer the same data file format for all splits. Got {NamedSplit('train'): (None, {}), NamedSplit('validation'): ('json', {})}
Error code:   FileFormatMismatchBetweenSplitsError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

OpenAI4S retrosynthesis scenarios

This public snapshot contains six independently specified retrosynthesis and reaction-prediction Scenarios, their frozen OpenAI4S query prompts, matched GT reference codebases, synthetic protocol fixtures, private-evaluator wiring, model deployment adapters, documentation, and offline tests.

Download

  • openai4s-retrosynthesis-complete-20260901.zip
  • Verify it with the digest in SHA256SUMS before extraction.

Included Scenario pairs

  1. reaction-class-unknown single-step retrosynthesis;
  2. budgeted multistep route planning;
  3. reaction atom mapping;
  4. forward product prediction;
  5. categorical reaction-condition recommendation;
  6. OOD reaction-yield estimation.

For every Scenario, the ZIP separates:

  • scenarios/queries/: the frozen input prompt, including descriptions of all public input files and fields;
  • scenarios/gt_codebases/: repository-reviewed reference entry points;
  • scenarios/openai4s_codebases/: the actual-generation boundary and provenance manifest;
  • scenarios/test_cases/: CC0 synthetic protocol cases, installer, and evaluator.

Important scientific boundary

The bundled JSON cases are synthetic protocol smoke tests. They validate file schemas, public/private Ground Truth isolation, deterministic artifacts, budgets, hashes, and evaluator wiring; they do not establish model accuracy.

Production USPTO, PaRoutes, atom-mapping, Parrot condition, and Buchwald--Hartwig HTE artifacts are not redistributed. Their registry entries remain not_frozen until a maintainer has reviewed the source revision, license, split, and file hashes.

The snapshot records an attempted real OpenAI4S CLI generation that stopped before the first model call because the configured Ark provider had no API key. No GT source was copied into the generated-code directory. Run the included generation driver after configuring an authorized provider to produce and cross-check the six independent OpenAI4S codebases.

Mirrored checkpoints and data

The following third-party artifacts are mirrored with their upstream model cards, licenses, fixed revisions, and hashes:

  • checkpoints/reactiont5v2-forward/: MIT snapshot of sagawa/ReactionT5v2-forward at revision 933114058cb2604dc1bf536dbebdfcefbe83d4fc;
  • checkpoints/reactiont5v2-yield-quarantined/: MIT snapshot of sagawa/ReactionT5v2-yield at revision f0658bfd360bceaaf560f11b850781c50221fe0b; this checkpoint is quarantined because the local public canary did not reproduce the model-card value;
  • checkpoints/parrot/: the two admitted MIT Parrot artifacts copied from xiaoruiwang/ChemEnzyRetroPlanner_metadata revision b9ef6049d341bfc62d835f09ad6ce33b6f86b047;
  • checkpoints/rxnmapper/: official MIT RXNMapper 0.4.3 PyPI wheel;
  • datasets/buchwald-hartwig-ood/: Apache-2.0 official repository snapshot at commit 3a44e398d9fed3f08cf396f1153be2960193f581 including the curated CSV;
  • datasets/paroutes-source/: Apache-2.0 PaRoutes source snapshot at commit bc5e69568092246f7c7bb5405216f8a3a334adbc.

The PaRoutes Zenodo route/model files are deliberately not mirrored because the record page does not expose an artifact license. AiZynthFinder public-data artifacts and processed USPTO datasets are also not mirrored. The MIT RetroChimera USPTO-50K checkpoint is registered, but its official Figshare host returned HTTP 403 from the upload environment; no unreviewed third-party mirror was substituted. See THIRD_PARTY_ARTIFACTS.json and THIRD_PARTY_NOTICES.md.

License and provenance

The synthetic Scenario fixtures are marked CC0-1.0. Source code and documents retain the licensing terms of their originating OpenAI4S repository. External datasets, model repositories, and checkpoints retain their own licenses and are not covered by the fixture license. This repository therefore uses the aggregate other card label rather than implying one license covers every referenced resource.


中文说明

本快照包含六个可独立运行和评测的逆合成/反应预测 Scenario、冻结的 OpenAI4S 输入 prompt、同名 GT 参考实现、合成协议测试数据、私有 evaluator、模型部署 adapter、 说明文档和离线测试。

包内测试数据只用于验证 schema、公开/私有 Ground Truth 隔离、预算、哈希和 evaluator 接线,不代表真实模型精度。真实 USPTO、PaRoutes、Parrot、原子映射和 Buchwald--Hartwig 数据及大型 checkpoint 没有重新分发。

真实 OpenAI4S CLI 首次生成尝试在模型调用前因 Ark API key 未配置而停止,状态已在 manifest 中如实记录;包内没有用 GT 源码冒充 OpenAI4S 生成结果。

仓库另外镜像了固定 revision 的 ReactionT5v2-forward、隔离状态的 ReactionT5v2-yield、Parrot、RXNMapper,以及 Apache-2.0 的 Buchwald–Hartwig OOD CSV。PaRoutes 只镜像 Apache-2.0 源码和下载协议,其 Zenodo 数据页未展示 artifact 许可证,因此没有二次分发路线文件。AiZynthFinder public-data、加工 USPTO 数据和 下载受阻的 RetroChimera checkpoint 也没有使用未经审查的替代来源。

Downloads last month
36