reasoning-dataset-lab Handbook

reasoning-dataset-lab (rdlab) builds datasets of puzzles, quizzes, verification items and reasoning items from pre-registered maps, bound canon and frozen parts, with provenance on every item. Its one rule is that the answer key comes from code, not from a model: code builds the item from a hidden formal state and proves the key, and a model is asked only to put the item into words.

Where to go

• Getting started: install it and validate a map with no model at all.
• The map-first principle: the map, cells and capacity, and the project fingerprint.
• Pilot B builds: the part that runs real local models: the memory plan, the models, calibration.
• Certification: how the generator’s value sets are certified, and what the v3 run found.
• Reference: the commands, exit codes and error codes.
• Security and data: what it reads and writes, and how to report a problem.

Status

Alpha, at 0.5.0, which adds the judge bake-off and its Colab launcher. Interfaces and data shapes can change without notice. Nothing rdlab writes is training data until a person has checked it.