melvil

Labeled examples + a label taxonomy in → an optimized, versioned classifier prompt out.

melvil wraps the GEPA reflective prompt-evolution engine with a classification-specific layer that generic prompt optimizers lack: confusion-driven reflection, a per-label prompt codebook, and hard-example mining. It is a pure Python library — the API is the product.

Named for Melvil Dewey, who gave libraries a system for putting things in the right category.

import melvil as mv

examples = mv.load_csv("tickets.csv")                  # text,label columns
train, dev = mv.train_dev_split(examples, dev_size=100, seed=0)
spec = mv.TaskSpec.from_examples("ticket-intents", examples)
cfg = mv.Config(task_model="openai/gpt-4.1-mini",
                reflection_model="openai/gpt-4.1", budget="light")
artifact = mv.optimize(spec, train, dev, cfg)
print(artifact.render())                               # deployable prompt string
artifact.save("ticket_intents.v1.json")

Install

pip install pymelvil             # the PyPI distribution is `pymelvil`; you `import melvil`
pip install 'pymelvil[hf]'       # + HuggingFace dataset loaders

From a checkout:

pip install -e '.[hf,dev]'       # + test/lint tooling

Model names are LiteLLM ids (openai/..., anthropic/..., openrouter/...); set the matching API key env var (OPENAI_API_KEY, OPENROUTER_API_KEY, ...).

What the benchmarks say (read this before choosing features)

We benchmark honestly, including against ourselves — full protocol and numbers in benchmarks/RESULTS.md. The pre-registered confirmation pass (8 public datasets, fresh seeds, light budget, gpt-4.1-mini) found:

• Vanilla GEPA over melvil's rendered prompt is the strongest configuration (mean test accuracy 0.781 vs 0.755 for the full classification layer and 0.762 for MIPROv2; seed prompt 0.703). At light budgets, prefer features=mv.Features.none().
• The classification layer's per-component updates trade whole-prompt coverage for structure; at ~6–10 accepted proposals per run that trade loses, especially on hard small-taxonomy tasks. Whether it wins at medium/heavy budgets is an open question.
• Hard-example mining, behind its (strict, quarantined) accept gate, kept exemplars in 0/24 confirmation runs — treat it as a safety-gated no-op at light budgets.

What the classification layer does

Each feature is independently toggleable via mv.Features (all off by default — the default is the benchmark-strongest vanilla-GEPA configuration; enable the layer with mv.Features.all() or per-flag; Features.none() is an explicit alias for the default):

1. Per-label codebook — the prompt is not a free-text blob but named components: a task instruction, one definition per label, and boundary rules. GEPA evolves them per-component.
2. Confusion-driven reflection — every full dev evaluation updates a confusion matrix; reflection rounds are pointed at the label components on both sides of the currently worst confused boundary, and the reflection LM is shown the top confused pairs with concrete misclassified examples.
3. Hard-example mining — dev examples that stay misclassified across the run become candidate few-shot exemplars, selected to cover the top confused boundaries, and kept only if they don't hurt dev accuracy. (Exemplars come from dev, so dev scores of exemplar-augmented artifacts are mildly optimistic — judge them on test.)

The rendered prompt always ends with a fixed, non-evolvable output-format contract, so the optimizer can never break parseability.

Everything else you get

• PromptArtifact — versioned JSON: components, models, budget and cost actually spent, dev scores + confusion matrix, the dev-score-vs-budget curve, config hash, lineage (parent_id). artifact.diff(other) gives a per-component diff with score deltas.
• evaluate(artifact, data, model=...) — accuracy, macro-F1, per-label P/R/F1, confusion, cost; pass a different model for a transfer evaluation. report(...) renders it as markdown.
• Cost estimation before spending — estimate_optimize_cost(spec, train, dev, cfg) is a dry-run upper bound; measured spend comes from the LM call history and lands in the artifact.
• Run directories & resume — every run writes runs/<task>/<hash>-s<seed>/ (config, engine state, reflection traces, artifact); optimize(..., resume=True) continues an interrupted run.
• Progress — pass on_round=lambda info: ... for live dev score / spend after every full dev evaluation, or just read the default logging.
• Offline testing — melvil.testing ships fake LMs with a real optimization gradient; the whole test suite runs with no API keys.

Worked example & docs

• examples/quickstart.py — the snippet above, runnable end-to-end on AG News.
• examples/agnews_demo.ipynb — notebook walkthrough with a live progress callback.
• docs/api.md — public API reference.
• benchmarks/ — the honest-benchmark harness (datasets × arms × seeds).

Development

pip install -e '.[dev]'
pytest          # green with no API keys — fake-LM offline suite
ruff check .

License: MIT.