PTTREP/asynchow-code-aligned-minutes

AsynChow Code-Aligned Minutes

This dataset is a unit-normalized variant of the AsynChow data released with `fangru-lin/procedure_generalization_llm`, pinned to source commit d9bf3485cd41c1050d33471d922c826f474efec1.

It contains three aligned representations of each weighted DAG scheduling problem:

• —natural: natural-language steps and precedence constraints;
• —graph: adjacency-list and duration-dictionary representation;
• —python: executable-style Python representation from the released Code data.

Each configuration contains 1,373 training examples and 225 test examples. Rows remain aligned across configurations, and source_row identifies the original row within each split.

Unit-alignment policy

• —Durations in the Natural and Graph inputs are converted to minutes.
• —Values use the same two-decimal rounding convention as the released Code representation.
• —All three representations use the released Code numeric target.
• —A month is treated as 30 days and a year as 365 days, matching the released Code generation rule.
• —The upstream source files are not modified.

Schema

Each JSONL row contains:

• —source_row: zero-based row number in the upstream split;
• —question: the aligned problem prompt;
• —answer: the numeric answer in minutes;
• —source_answer: the answer from the corresponding upstream representation.

Known rounding differences

The Code representation's two-decimal convention introduces quantization. The Code target differs from both exact Natural/Graph interval endpoints for 136 of 1,373 training examples and 16 of 225 test examples. It falls outside the released interval for 97 training examples and 12 test examples. These differences are intentionally retained to keep all representations aligned to the Code target.

See manifest.json for the complete conversion policy, source and generated SHA-256 hashes, and audit counts. The conversion script is included as prepare_code_aligned_minutes.py.

License and attribution

This repository is a transformed redistribution of the upstream AsynChow data. No explicit dataset license was available in the local pinned source materials when this variant was prepared. Users should review the upstream repository and obtain any permissions required for their intended use.