Tune.Optimizers.SIMBA (Tune v0.5.0)

Copy Markdown View Source

Stochastic introspective mini-batch ascent.

Each of max_steps steps:

  1. takes the next training minibatch of bsz rows from a seeded shuffle;
  2. runs num_candidates trajectories per row, each with a program sampled from the pool by a softmax over mean minibatch scores (temperature);
  3. ranks rows by the gap between their best and worst trajectory scores, then by lowest best score, so variable and then weak behavior is targeted first;
  4. for the top num_candidates rows, builds a new candidate from a sampled pool program with one strategy, chosen at random:
    • append_a_demo adds the best trajectory's final call per predictor as a demonstration (keeping at most max_demos),
    • append_a_rule asks the proposal role to contrast the best and worst trajectory and appends the advice to one predictor's instruction; the positive trajectory is the row's own best when it succeeded and beat its worst; for a weak row with no gap, it is the best successful trajectory of another minibatch row that scored higher. A row with no such positive is skipped, so failures never become demonstrations;
  5. scores new candidates on the same minibatch and adds them to the pool.

Finally the baseline and the num_final best pool programs by mean minibatch score are evaluated on the full validation set; only a complete, eligible validation result is returned.

Options: :bsz (default 4), :num_candidates (default 4), :max_steps (default 4), :max_demos (default 4), :temperature (default 0.2), :num_final (default 3).

Differences from DSPy: trajectory variation comes from sampling pool programs, not from changing the model temperature; final selection uses the explicit validation set.