Prompt Regression

from openjudge.graders.llm_grader import LLMGrader
from openjudge.graders.schema import GraderMode
from openjudge.runner.grading_runner import GradingRunner
from openjudge.analyzer.pairwise_analyzer import PairwiseAnalyzer

# Judge prompt for relevance comparison
relevance_judge = LLMGrader(
    model=model,
    name="relevance_compare",
    mode=GraderMode.POINTWISE,
    template="""
Compare Response A and Response B for the query below.
Which response better addresses the user's question?

Query: {query}
Response A: {response_a}
Response B: {response_b}

Score 1.0 if A is better, 0.0 if B is better, 0.5 if tied.
Respond in JSON: {{"score": <float>, "reason": "<explanation>"}}
""",
)

# Build pairwise dataset with position swap
dataset = []
for sample in test_samples:
    # Original order
    dataset.append({
        "query": sample["query"],
        "response_a": baseline_outputs[sample["id"]],
        "response_b": candidate_outputs[sample["id"]],
        "metadata": {"model_a": "baseline", "model_b": "candidate"},
    })
    # Swapped order — critical for debiasing
    dataset.append({
        "query": sample["query"],
        "response_a": candidate_outputs[sample["id"]],
        "response_b": baseline_outputs[sample["id"]],
        "metadata": {"model_a": "candidate", "model_b": "baseline"},
    })

runner = GradingRunner(
    grader_configs={"relevance": relevance_judge},
    max_concurrency=8,
)
results = await runner.arun(dataset)

# Analyze with PairwiseAnalyzer
analyzer = PairwiseAnalyzer(model_names=["baseline", "candidate"])
analysis = analyzer.analyze(dataset, results["relevance"])

print(f"Win rates: {analysis.win_rates}")
# → {'baseline': 0.35, 'candidate': 0.55} → candidate wins 55% of comparisons
print(f"Best model: {analysis.best_model}")

import numpy as np
from openjudge.graders.schema import GraderScore

def per_comparison_winners(dataset, grader_results):
    """One named winner per comparison row (handles swapped order via metadata)."""
    winners = []
    for sample, result in zip(dataset, grader_results):
        if not isinstance(result, GraderScore):
            continue  # skip errors
        meta = sample.get("metadata", {})
        winners.append(meta["model_a"] if result.score >= 0.5 else meta["model_b"])
    return winners

def bootstrap_win_rate(winners, target, n_iter=1000):
    n = len(winners)
    rates = []
    for _ in range(n_iter):
        idx = np.random.choice(n, n, replace=True)
        rates.append(sum(1 for i in idx if winners[i] == target) / n)
    return float(np.percentile(rates, 2.5)), float(np.percentile(rates, 97.5))

winners = per_comparison_winners(dataset, results["relevance"])
n = len(winners)
candidate_rate = sum(1 for w in winners if w == "candidate") / n
baseline_rate = sum(1 for w in winners if w == "baseline") / n
ci_low, ci_high = bootstrap_win_rate(winners, target="candidate")

if ci_low > 0.5:
    verdict = "candidate BETTER"
elif ci_high < 0.5:
    verdict = "candidate WORSE"
elif (ci_high - ci_low) < 0.3:
    verdict = "TIED (CI brackets 0.5, narrow)"
else:
    verdict = "INCONCLUSIVE (CI too wide — need more samples)"

print({"candidate_win_rate": candidate_rate, "baseline_win_rate": baseline_rate,
       "ci_95": [ci_low, ci_high], "verdict": verdict})

Prompt Regression: v1 (baseline) vs v2 (candidate)
Task: Customer support chatbot
Samples: 50

Dimension            Candidate  Baseline  Tie   95% CI         Verdict
===========================================================================
Answer relevance        58%       32%      10%   [51%, 65%]   ✓ BETTER
Factual accuracy        48%       44%       8%   [41%, 55%]   = TIED
Tone appropriateness    38%       52%      10%   [31%, 45%]   ✗ WORSE
Conciseness             62%       28%      10%   [55%, 69%]   ✓ BETTER

Summary: v2 is significantly better on relevance and conciseness,
but worse on tone appropriateness. The tone regression likely comes
from the new "be direct" instruction — consider softening it.

Top 3 tone failures (candidate worse):
  1. Query: "I'm really frustrated..." → v2 response too curt
  2. Query: "This is my first time..." → v2 missing empathetic opening
  3. Query: "Can you help me understand..." → v2 skipped explanation