AutoMetrics

Automatically induce evaluation metrics that approximate human judgment from fewer than 100 labels.

AutoMetrics takes a small set of human-labeled examples (thumbs, Likert, pairwise — under 100 points) and produces a single interpretable evaluator for your task. It synthesizes candidate criteria with LLM judges, retrieves complementary metrics from a curated bank of 48, and composes them with PLS regression. Across five tasks in the paper, it beats LLM-as-a-judge baselines by up to +33.4% Kendall τ, and in an agentic-task case study it matches the performance of a verifiable reward.

---

Install

pip install autometrics-ai

Base install requires only Python 3.9+. Heavy dependencies (Java 21, pyserini, pylate, bert_score, …) are loaded lazily — they're needed only if you opt into features that use them.

Quickstart

export OPENAI_API_KEY="sk-..."
python examples/tutorial.py

Builds a tiny custom dataset, generates a handful of LLM-judge metrics for your task, fits PLS to your human scores, and writes an interactive HTML report to artifacts/. No Java, no GPU, no bank dependencies required for this path.

How it works

1. Generate. Propose task-specific candidate metrics — single-criterion, rubric, example-based, and MIPROv2-optimized LLM judges (10 + 5 + 1 + 1 by default).
2. Retrieve. Rank the generated candidates alongside the 48-metric MetricBank (ColBERT → LLM reranker) and keep the top k=30.
3. Regress. Fit Partial Least Squares on the training set to select n=5 predictive metrics and learn their weights.
4. Report. Emit (a) the aggregated metric as a Python class you can import, (b) a Metric Card per generated metric, and (c) an HTML report card with coefficients, correlation, robustness, runtime, and per-example feedback.

For datasets of ≤100 rows AutoMetrics runs in generated-only mode by default, skipping the metric bank entirely.

See the paper (ICLR 2026) for the full method, ablations, and case study.

Examples

File	Scope	Requires
examples/tutorial.py	Dead-simple 8-row demo, generated-only	OPENAI_API_KEY
examples/autometrics_simple_example.py	Full pipeline with defaults on HelpSteer	+ Java 21, bank extras
examples/autometrics_example.py	Custom generators, retriever, regressor, priors	+ your own config

Narrative walkthrough: examples/TUTORIAL.md.

Use on your own data

import dspy, pandas as pd
from autometrics.autometrics import Autometrics
from autometrics.dataset.Dataset import Dataset

df = pd.DataFrame({
    "id": ["1", "2", "3"],
    "input":  ["prompt 1", "prompt 2", "prompt 3"],
    "output": ["response 1", "response 2", "response 3"],
    "score":  [4.5, 3.2, 4.8],
})
dataset = Dataset(
    dataframe=df, name="MyTask",
    data_id_column="id", input_column="input", output_column="output",
    target_columns=["score"], ignore_columns=["id"], metric_columns=[],
    reference_columns=[], task_description="Describe your task in one sentence.",
)

llm = dspy.LM("openai/gpt-4o-mini")
results = Autometrics().run(
    dataset=dataset, target_measure="score",
    generator_llm=llm, judge_llm=llm,
)

final = results["regression_metric"]       # an importable Metric
final.predict(dataset)                     # scores on any Dataset with same schema

Requirements

Component	Needed for
Python ≥ 3.9	everything
OPENAI_API_KEY (or any LiteLLM-compatible endpoint)	LLM-based generation and judging
Java 21	BM25 retrieval over the full MetricBank (pyserini)
GPU	some bank metrics (reward models, large BERTScore); CPU works for generated-only

Repository layout

autometrics/
├── autometrics.py            Pipeline orchestrator
├── dataset/                  Dataset interface + built-in tasks
├── metrics/                  MetricBank (48 metrics) + generated metric scaffolds
├── generator/                LLM judge proposers (single, rubric, examples, G-Eval, optimized)
├── recommend/                Retrievers (BM25, ColBERT, LLMRec, Pipelined)
├── aggregator/regression/    PLS (default), Lasso, Ridge, ElasticNet, HotellingPLS
└── util/report_card.py       HTML report generator
examples/                     Tutorial scripts and walkthroughs

Optional extras

Install extras for metric-bank components with heavier dependencies

pip install "autometrics-ai[bert-score,rouge,bleurt]"
pip install "autometrics-ai[reward-models,gpu]"
pip install "autometrics-ai[mauve,parascore,lens,fasttext]"

Individual clusters: fasttext, lens, parascore, bert-score, bleurt, moverscore, rouge, meteor, infolm, mauve, spacy, hf-evaluate, reward-models, readability, gpu. See pyproject.toml for the full mapping. Metrics whose dependencies are missing are silently skipped with a warning — no install is strictly required.

Citation

@inproceedings{ryan2026autometrics,
  title   = {AutoMetrics: Approximate Human Judgments with Automatically Generated Evaluators},
  author  = {Ryan, Michael J and Zhang, Yanzhe and Salunkhe, Amol and Chu, Yi and Xu, Di and Yang, Diyi},
  booktitle = {The Fourteenth International Conference on Learning Representations},
  year    = {2026},
  url     = {https://openreview.net/forum?id=ymJuBifPUy}
}

License

MIT — see LICENSE.