Optimizers

Optimizers compile a program against labeled data and a metric, producing a new program with tuned demos and/or instructions.

Few-shot compilation

Compile demos from labeled data:

{:ok, compiled, _stats} =
  Dsxir.compile(
    Dsxir.Optimizer.BootstrapFewShot,
    prog,
    trainset,
    &MyApp.Metric.f1/3,
    max_bootstrapped_demos: 4
  )

Dsxir.save!(compiled, "qa.v1.json")

Dsxir.Optimizer.LabeledFewShot (no LM calls) and Dsxir.Optimizer.BootstrapFewShot (trace-driven, with diversity) are the two v0 few-shot optimizers.

MIPROv2

Dsxir.Optimizer.MIPROv2 jointly searches over candidate instructions and demo bundles. It bootstraps demo candidates, asks a proposer LM for candidate instructions grounded in program and dataset summaries, then runs a sampler-driven search (TPE by default) with periodic full-valset reranks of the top trials.

{:ok, compiled, stats} =
  Dsxir.compile(
    Dsxir.Optimizer.MIPROv2,
    program,
    trainset,
    &MyApp.Metric.f1/3,
    auto: :medium
  )

stats.best_score

auto: accepts :light | :medium | :heavy (see Dsxir.Optimizer.MIPROv2.Auto). Notable overrides: :proposer_lm, :sampler, :batch_size, :seed, :minibatch_full_eval_steps, :top_k_full_eval.

COPRO

Dsxir.Optimizer.COPRO is an instruction-only prompt optimizer that tunes each predictor's instruction independently by greedy coordinate ascent. For each round (up to depth rounds), it generates breadth candidate instructions per predictor, scores each against the full training set, and keeps the best. It does not touch demos or few-shot examples, making it cheaper than MIPROv2 and a good warm-up step before a full joint search.

{:ok, compiled, stats} =
  Dsxir.compile(
    Dsxir.Optimizer.COPRO,
    program,
    trainset,
    &MyApp.Metric.f1/3,
    auto: :medium
  )

stats.best_score

auto: accepts :light | :medium | :heavy (see Dsxir.Optimizer.COPRO.Auto). Preset values — light: breadth 4 / depth 2, medium: breadth 6 / depth 3, heavy: breadth 10 / depth 4. Notable overrides: :breadth, :depth, :init_temperature.

SIMBA

Dsxir.Optimizer.SIMBA (Stochastic Introspective Mini-Batch Ascent) searches over demos and instructions via mini-batch trajectory sampling and variance-driven bucketing. Each step softmax-picks source programs from a growing pool, runs diverse executions per mini-batch example, and applies one of two strategies: AppendDemo adds the best I/O trace as augmented demos, and AppendRule asks a reflective LM to contrast better-vs-worse trajectory pairs and appends per-predictor advice to each instruction. A post-loop finalize scores evenly-spaced winners on the full trainset and returns the best.

Reach for SIMBA when you want both demo augmentation and self-reflective rule generation in one mini-batch ascent pass. Prefer MIPROv2 for a sampler-driven joint search over a large instruction proposal set, GEPA for a reflective proposal search without the stochastic mini-batch loop, or COPRO when only instructions need tuning.

{:ok, compiled, stats} =
  Dsxir.compile(
    Dsxir.Optimizer.SIMBA,
    program,
    trainset,
    &MyApp.Metric.f1/3,
    auto: :medium
  )

stats.best_score

auto: accepts :light | :medium | :heavy (see Dsxir.Optimizer.SIMBA.Auto).

:auto preset values:

SIMBA implements all four checkpointable Dsxir.Optimizer callbacks and works with Dsxir.OptimizerSession for pause/resume across steps.