| ---
|
| license: cc-by-4.0
|
| language: [en]
|
| task_categories: [text-generation]
|
| tags: [code, dsl, business-process, safety]
|
| pretty_name: Aether synthetic SOP-to-program dataset
|
| ---
|
|
|
| # Aether synthetic SOP-to-program dataset
|
|
|
| Pairs of plain-English business procedures (SOPs) and programs in Aether, a small language for checkable business procedures. Each row also carries the owner's policy (safety rules) and the sheet declarations the program may use.
|
|
|
| **How it was made:** a rule-based generator (no language model) builds the SOP, the policy, and the program from the same random choices, so they agree by construction. A row was kept only if its program parses, passes the Aether checker with zero errors, and runs on 3 random mock workspaces with no run error and no monitor violation.
|
|
|
| | Split | Rows |
|
| |---|---|
|
| | train | 1649 |
|
| | valid | 125 |
|
| | test | 392 (held-out business types: florist, photostudio, plumbing, vet) |
|
|
|
| 2166 rows in total, from 10 procedure patterns.
|
|
|
| **Fields:** `id`, `domain`, `pattern`, `batch` (approval mode), `sop`, `policy`, `sheets`, `program`.
|
|
|
| **Trap-style samples** (`traps_train.jsonl`, `traps_valid.jsonl`; used to train model v0.2): the SOP asks for something the policy forbids or Aether cannot do, and the program does the permitted alternative. These samples were checked for overlap with every benchmark and held-out SOP; see `traps_overlap_report.json`. They have an extra field, `trap`. Training on them did not improve held-out trap performance (see the model card).
|
|
|
| **Limitations:** the English is templated, so it is far less varied than real owners' writing. All businesses are invented, and all data in the programs is mock data.
|
|
|
