Tune.Optimizers.Avatar (Tune v0.5.0)

Copy Markdown View Source

Contrastive instruction improvement for one named predictor.

The unchanged program is evaluated on validation first. Then, for up to max_iters iterations, the current program runs on the training set. Trajectories scoring at least upper_bound are positive; failed runs and scores at most lower_bound are negative, so a failure is never positive. The proposal role compares up to max_positive positive and max_negative negative trajectories and writes feedback, then rewrites the named predictor's instruction from that feedback. Each rewrite is validated before it can be returned; only that predictor's instruction changes. If either group is empty the search stops cleanly with stop note :no_contrast and returns the best validated program.

Options: :predictor (required name), :upper_bound and :lower_bound (required scores), :max_iters (default 3), :max_positive and :max_negative (default 5).

Differences from DSPy: this operates on a declared predictor instruction and ordinary program traces; it needs no Avatar agent runtime or tools.