TrainRation

Decide what to train on, and never move the yardstick.

Data rationing and immutable evaluation yardsticks for robot learning.

Status: early, pre-alpha

Nothing here yet. This organization is reserved for an in-development project; datasets and models will appear as the work becomes reproducible rather than merely runnable.

Current artifacts live elsewhere:

Source — github.com/trainration/trainration
PyPI — trainration (0.0.1, name reservation, no functional API)

What it will do

Robot learning data is expensive to collect and much of it is redundant. TrainRation works on two problems:

Deciding what to train on — dataset selection, mixture ratios, curriculum ordering, and failure attribution over LeRobot v3.0 datasets.
Keeping the yardstick fixed — the evaluation criteria that decide whether a change helped must not be reachable by the process that proposes changes. A self-improving system that can edit its own success criteria will optimize the criteria instead of the policy.

The second point is the design constraint that shapes the first. Research direction evolves; the measuring stick does not.

Planned components

Package	Scope
`trainration-probe`	Cross-dataset quality and consistency checks for LeRobot v3.0
`trainration-yardstick`	Evaluation criteria and admission constraints as structural, non-bypassable boundaries
`trainration-core`	Recursive decision kernel: mixture ratios, failure attribution, curriculum

What will land here

This org is the home for the data-side artifacts:

Curated LeRobot v3.0 dataset subsets, with the selection criteria published alongside them
Held-out evaluation splits that are versioned and immutable once released
Probe reports over public robot learning datasets

License: MIT