cje-eval 0.2.4

Causal Judge Evaluation - Unbiased LLM evaluation framework

CJE - Causal Judge Evaluation

Turn noisy LLM-judge scores into precise, unbiased estimates of the outcomes you care about.

CJE calibrates judge scores using a small oracle slice (5-10% coverage), then delivers statistically rigorous estimates.

How It Works

CJE follows a simple three-step workflow:

┌─────────────────────────────────┐
│           Data                  │
│  LLM-judge scores +             │
│  oracle slice (5-50%)           │
└─────────────────────────────────┘
              ↓
┌─────────────────────────────────┐
│         Calibrate               │
│  Learn judge → oracle mapping   │
└─────────────────────────────────┘
              ↓
┌─────────────────────────────────┐
│          Estimate               │
│  Estimates with honest          │
│  uncertainty                    │
└─────────────────────────────────┘

Key benefits:

• Small label budget: 5-10% oracle coverage often sufficient
• Unbiased estimates: Judge scores (+ optional covariates) mapped to oracle scale
• Rigorous inference: CIs account for both sampling and calibration uncertainty

See cje/calibration/README.md for technical details.

📊 Performance

Arena Experiment: 5k Real Evaluations - Comprehensive benchmarking on ChatBot Arena data:

• 94% pairwise ranking accuracy with Direct Model + covariates
• 158× ESS improvement with SIMCal-W vs raw SNIPS
• Kendall τ = 0.837 vs -0.235 for uncalibrated methods
• Validates AutoCal-R calibration and doubly-robust estimation on real data

Reproduction code: Full experimental pipeline available at cje-arena-experiments

Calibration Methods

CJE provides two calibration modes for mapping judge scores to oracle outcomes (i.e., the KPI you care about):

Monotone

Standard isotonic regression enforces: higher judge score → no worse expected outcome.

Why isotonic? It's the right structural prior—assumes only monotonicity (which you actually believe), preserves oracle KPI levels by construction (mean-preserving by KKT conditions), and is highly efficient with small label budgets. See technical rationale.

Simple, stable, works well when the judge-oracle relationship is already monotone.

Two-Stage (Default with Covariates)

Learns smooth transformation g(S, X) → rank → isotonic. Handles non-monotone patterns and incorporates additional covariates (e.g., response length, domain metadata) while maintaining final monotonicity guarantee.

_{Two-stage calibration learns flexible relationships between covariates (judge score, response length) and oracle outcomes in Stage 1, then enforces monotonicity via isotonic regression in Stage 2. Left/Middle: Partial dependence plots show how each covariate relates to oracle score while holding others at mean values. Right: Final monotone mapping from Stage 1 risk index to calibrated oracle score. Full benchmarking results: Arena Experiment. Data from LMSYS Chatbot Arena.}

When to use two-stage:

• You have covariates (response length, domain, etc.) → two-stage is default and recommended
• Judge shows non-monotone empirical E[Oracle|Judge] relationship
• Regional miscalibration (monotone works well at low/high but poorly at mid-range)
• Length bias (judge gives different scores to same-quality responses based on length)

Auto mode:

• With covariates: Two-stage is automatically used (can incorporate additional features)
• Judge score only: CJE automatically selects monotone vs two-stage via cross-validation (1-SE rule)

# Default: Judge score only (no covariates, auto-selects monotone/two-stage via CV)
result = analyze_dataset(fresh_draws_dir="responses/")

# Include response_length covariate for two-stage calibration
result = analyze_dataset(
    fresh_draws_dir="responses/",
    include_response_length=True
)

# Add domain as additional covariate (combine with response_length)
result = analyze_dataset(
    fresh_draws_dir="responses/",
    include_response_length=True,
    calibration_covariates=["domain"]
)

# Force a specific mode
result = analyze_dataset(
    fresh_draws_dir="responses/",
    calibration_mode="monotone"  # or "two_stage"
)

Installation

pip install cje-eval

🚀 Try it Now - Interactive Demo

Quick Start

Minimal Example

from cje import analyze_dataset

# Compare policies on an eval set
result = analyze_dataset(fresh_draws_dir="responses/")

# Get estimates with confidence intervals
for policy, est, se in zip(
    result.metadata["target_policies"],
    result.estimates,
    result.standard_errors
):
    print(f"{policy}: {est:.3f} ± {1.96*se:.3f}")

Data Format

Directory structure: One JSONL file per policy

responses/
├── model_a_responses.jsonl
└── model_b_responses.jsonl

Minimal record (inside each file):

{"prompt_id": "eval_0", "judge_score": 0.85}
{"prompt_id": "eval_1", "judge_score": 0.72}

With calibration (add oracle labels to 5-10% of samples):

{"prompt_id": "eval_0", "judge_score": 0.85, "oracle_label": 0.86}
{"prompt_id": "eval_1", "judge_score": 0.72}

CJE automatically:

• Discovers policies from filenames (model_a_responses.jsonl → policy "model_a")
• Applies AutoCal-R when oracle labels are present
• Uses cluster-robust SEs for paired comparisons (when same prompts across policies)
• Returns unbiased estimates with valid 95% CIs

Paired Comparisons

When comparing policies on the same prompts (paired design), CJE automatically uses cluster-robust standard errors:

# Both files must have matching prompt_ids for pairing
result = analyze_dataset(fresh_draws_dir="responses/")

# CJE automatically clusters by prompt for valid inference
if result.metadata.get("prompts_aligned"):
    print("✓ Paired comparison - using cluster-robust SEs")

Why it matters: Paired designs have correlated outcomes across policies (same prompt evaluated by multiple models). Standard SEs would understate uncertainty. CJE automatically accounts for this by clustering by prompt_id.

Beyond Direct Mode

CJE also supports IPS (counterfactual inference from logs) and DR (doubly robust with fresh draws). These require log probabilities from your models.

# IPS: Estimate "what if we deployed policy X?" from existing logs
result = analyze_dataset(logged_data_path="logs.jsonl")

# DR: Combine logged data + fresh draws for maximum accuracy
result = analyze_dataset(
    logged_data_path="logs.jsonl",
    fresh_draws_dir="responses/"
)

For IPS/DR data formats and API details: Run help(analyze_dataset) or see cje/interface/ module docs.

Visualization

CJE provides diagnostic plots for understanding and validating results:

from cje import analyze_dataset, plot_policy_estimates

# Run analysis
result = analyze_dataset(fresh_draws_dir="responses/")

# Quick plot with convenience method
result.plot_estimates(save_path="estimates.png")

# Or use visualization functions directly for more control
plot_policy_estimates(
    estimates={"policy_a": 0.75, "policy_b": 0.68},
    standard_errors={"policy_a": 0.02, "policy_b": 0.03},
    oracle_values={"policy_a": 0.74, "policy_b": 0.69}  # Optional
)

Available visualizations:

• plot_policy_estimates - Forest plots with confidence intervals
• plot_calibration_comparison - Judge→oracle calibration curves
• plot_weight_dashboard_summary - Weight diagnostics for IPS/DR
• plot_weight_dashboard_detailed - Per-policy weight analysis
• plot_dr_dashboard - Doubly robust diagnostics

Jupyter notebooks: Results automatically display as formatted tables when evaluated in a cell.

See cje/visualization/README.md for complete guide.

Documentation

📚 Getting Started

• Interactive Demo - Try in your browser
• Examples - Working code samples

🔧 For Engineers

• Calibration Methods - AutoCal-R, isotonic regression, two-stage fallback
• Diagnostics System - Uncertainty quantification, OUA, transportability tests
• Estimators - Direct, IPS, DR implementations
• Interface/API - analyze_dataset implementation and mode selection

Development

git clone https://github.com/cimo-labs/cje.git
cd cje
poetry install
make test

Support

• 🐛 Issues
• 💬 Discussions

License

MIT - See LICENSE for details.