refua-bench

refua-bench is a standalone benchmark and regression-gating project for Refua model workflows. It benchmarks current and future models via adapter interfaces and enforces safe regression gates.

What It Provides

• Benchmark suite schema (yaml/json) for tasks, metrics, tolerances, and case sets.
• Adapter system for model execution:
  • golden: uses expected outputs (sanity checks)
  • file: reads predictions from a JSON artifact
  • command: calls any executable that reads JSON stdin and returns JSON stdout
  • custom adapters via module.path:AdapterClass
• Run artifacts in JSON + markdown.
• Automatic run provenance capture (git/runtime/model/dependencies).
• Statistical regression gating (minimum practical effect + bootstrap confidence intervals).
• Baseline registry with named baselines and safe promotion flow.

Install

cd refua-bench
poetry install

Build

poetry build

CLI

poetry run refua-bench --help

1. Run a benchmark

poetry run refua-bench run \
  --suite benchmarks/sample_suite.yaml \
  --adapter file \
  --adapter-config benchmarks/sample_file_adapter_config.yaml \
  --model-name boltz2-affinity \
  --model-version 2026-02-12 \
  --output artifacts/candidate_run.json \
  --markdown artifacts/candidate_run.md

By default, each run stores provenance in run.provenance.

2. Compare candidate vs baseline

poetry run refua-bench compare \
  --suite benchmarks/sample_suite.yaml \
  --baseline benchmarks/sample_baseline_run.json \
  --candidate artifacts/candidate_run.json \
  --output artifacts/compare.json \
  --markdown artifacts/compare.md

3. Statistical gating

poetry run refua-bench compare \
  --suite benchmarks/sample_suite.yaml \
  --baseline benchmarks/sample_baseline_run.json \
  --candidate artifacts/candidate_run.json \
  --output artifacts/compare_stats.json \
  --min-effect-size 0.02 \
  --bootstrap-resamples 2000 \
  --confidence-level 0.95 \
  --bootstrap-seed 7 \
  --fail-on-uncertain

Interpretation:

• min-effect-size: ignores changes too small to matter practically.
• bootstrap-resamples: enables CI-based robustness checks.
• --fail-on-uncertain: optional strict mode for inconclusive bootstrap tasks.

4. Run + compare in one command (gate)

poetry run refua-bench gate \
  --suite benchmarks/sample_suite.yaml \
  --baseline benchmarks/sample_baseline_run.json \
  --adapter file \
  --adapter-config benchmarks/sample_file_adapter_config.yaml \
  --candidate-output artifacts/candidate_run.json \
  --output artifacts/gate_report.json \
  --min-effect-size 0.02 \
  --bootstrap-resamples 1000

5. Baseline registry and promotion

Promote an initial baseline:

poetry run refua-bench baseline promote \
  --registry artifacts/baseline_registry.json \
  --suite benchmarks/sample_suite.yaml \
  --baseline-name stable \
  --candidate benchmarks/sample_baseline_run.json

Compare against named baseline:

poetry run refua-bench compare \
  --suite benchmarks/sample_suite.yaml \
  --registry artifacts/baseline_registry.json \
  --baseline-name stable \
  --candidate artifacts/candidate_run.json \
  --output artifacts/compare_named.json

Promote a new candidate safely (fails if regression is detected):

poetry run refua-bench baseline promote \
  --registry artifacts/baseline_registry.json \
  --suite benchmarks/sample_suite.yaml \
  --baseline-name stable \
  --candidate artifacts/candidate_run.json \
  --min-effect-size 0.02 \
  --bootstrap-resamples 2000

List/resolve baselines:

poetry run refua-bench baseline list --registry artifacts/baseline_registry.json
poetry run refua-bench baseline resolve \
  --registry artifacts/baseline_registry.json \
  --suite benchmarks/sample_suite.yaml \
  --baseline-name stable

6. Scaffold a new suite

poetry run refua-bench init --directory benchmarks/new_suite --name refua-next

Suite Schema

name: refua-core-smoke
version: 1.0.0
description: smoke checks
tasks:
  - id: affinity_mae
    metric: mae
    prediction_key: affinity
    expected_key: affinity  # optional, defaults to prediction_key
    regression_tolerance: 0.05
    weight: 2.0
    positive_label: 1       # used by f1/enrichment_factor/bedroc
    enrichment_fraction: 0.01  # used by enrichment_factor/ef
    bedroc_alpha: 20.0  # used by bedroc
    cases:
      - id: case_1
        input: {target: KRAS, ligand: MRTX1133}
        expected: {affinity: -9.3}

Supported metrics:

• mae
• rmse
• accuracy
• exact_match
• f1 (binary)
• enrichment_factor / ef (binary labels + ranking scores)
• bedroc (binary labels + ranking scores with early-recognition emphasis)

Prediction File Format (file adapter)

{
  "affinity_mae": {
    "case_1": {"affinity": -9.1}
  }
}

Command Adapter Contract

Input (stdin):

{
  "task_id": "affinity_mae",
  "prediction_key": "affinity",
  "case_id": "case_1",
  "input": {"target": "KRAS", "ligand": "MRTX1133"}
}

Output (stdout):

{"affinity": -9.2}

Tests

poetry run pytest