autometrics-research 0.0.6

Automatically generate evaluators that approximate human judgements

autometrics

AutoMetrics: Automatically discover, generate, and aggregate evaluation metrics for NLP tasks.

Autometrics helps you evaluate text generation systems by:

1. Generating task-specific candidate metrics with LLMs (LLM-as-a-judge, rubric/code generated metrics)
2. Retrieving the most relevant metrics from a bank of 40+ built-in metrics
3. Evaluating all metrics on your dataset (reference-free and reference-based)
4. Selecting the top metrics using regression
5. Aggregating them into a single, optimized metric and producing a report card

The repository includes simple scripts, examples, notebooks, and a full library to run the end-to-end pipeline.

Installation (pip + optional extras)

Install the published package (recommended):

pip install autometrics-research

Install with extras (examples):

pip install "autometrics-research[mauve]"
pip install "autometrics-research[bleurt,bert-score,rouge]"
pip install "autometrics-research[reward-models,gpu]"  # reward models + GPU accel

Developer install (from source):

pip install -e .

Optional extras (summary)

• fasttext: FastText classifiers — metrics: FastTextEducationalValue, FastTextToxicity, FastTextNSFW
• lens: LENS metrics — metrics: LENS, LENS_SALSA
• parascore: Paraphrase metrics — metrics: ParaScore, ParaScoreFree
• bert-score: metrics: BERTScore
• bleurt: metrics: BLEURT
• moverscore: metrics: MOVERScore (adds pyemd)
• rouge: metrics: ROUGE, UpdateROUGE
• meteor: metrics: METEOR (adds beautifulsoup4)
• infolm: metrics: InfoLM (adds torchmetrics)
• mauve: metrics: MAUVE (evaluate + mauve-text)
• spacy: metrics: SummaQA (requires spacy model; install with python -m spacy download en_core_web_sm)
• hf-evaluate: HF evaluate wrappers — metrics: Toxicity; also used by some wrappers
• reward-models: Large HF reward models — metrics: PRMRewardModel, INFORMRewardModel, LDLRewardModel, GRMRewardModel
• readability: metrics: FKGL (textstat)
• gpu: FlashAttention + NV libs (optional acceleration; benefits large reward models)

Quickstart

1. Install dependencies

pip install autometrics-research

2. Ensure Java 21 is installed (required by some retrieval components). See Java section below.

3. Set an API key for an OpenAI-compatible endpoint (for LLM-based generation/judging):

export OPENAI_API_KEY="your-api-key-here"

4. Run the simplest end-to-end example with sensible defaults:

python autometrics_simple_example.py

This will:

• load the HelpSteer dataset
• generate and retrieve metrics
• select top-k via regression
• print a summary and report card

For a power-user example with customization, run:

python autometrics_example.py

Examples and Tutorials

• Simple script with all defaults: examples/autometrics_simple_example.py
• Power-user/custom configuration: examples/autometrics_example.py
• Notebook tutorials: examples/tutorial.ipynb, demo.ipynb
• Text walkthrough tutorial: examples/TUTORIAL.md and runnable examples/tutorial.py

If you prefer an experiments-style entry point with CLI arguments, see:

python analysis/main_experiments/run_main_autometrics.py <dataset_name> <target_name> <seed> <output_dir>

There are also convenience scripts in analysis/ for ablations and scaling.

Repository Structure

• autometrics/dataset/datasets: Built-in datasets (e.g., helpsteer, simplification, evalgen, iclr, ...). The main dataset interface lives in autometrics/dataset/Dataset.py.
• autometrics/metrics: Metric implementations and utilities. See autometrics/metrics/README.md for how to write new metrics.
• autometrics/metrics/llm_judge: LLM-as-a-judge rubric generators (e.g., G-Eval, Prometheus-style, example-based).
• autometrics/aggregator/regression: Regression-based selection/aggregation (Lasso, Ridge, ElasticNet, PLS, etc.).
• autometrics/recommend: Metric retrieval modules (BM25/ColBERT/LLMRec and PipelinedRec).
• autometrics/test: Unit and integration tests, including caching behavior and generator tests.
• analysis/: Experiment drivers (CLI), ablations, robustness/scaling studies, and utilities.

Basic Usage (Library)

import os
import dspy
from autometrics.autometrics import Autometrics
from autometrics.dataset.datasets.helpsteer.helpsteer import HelpSteer

os.environ["OPENAI_API_KEY"] = "your-key-here"

dataset = HelpSteer()
generator_llm = dspy.LM("openai/gpt-4o-mini")
judge_llm = dspy.LM("openai/gpt-4o-mini")

autometrics = Autometrics()
results = autometrics.run(
    dataset=dataset,
    target_measure="helpfulness",
    generator_llm=generator_llm,
    judge_llm=judge_llm,
)

print([m.get_name() for m in results['top_metrics']])
print(results['regression_metric'].get_name())

For more advanced configuration (custom generators, retrieval pipelines, priors, parallelism), see TUTORIAL.md.

System Requirements

Python

Install dependencies:

pip install -r requirements.txt

Some metrics require GPUs. You can inspect GPU memory needs by checking gpu_mem on metric classes. Many metrics run on CPU.

Java (required for certain retrieval options)

This package requires Java Development Kit (JDK) 21 for some of its search functionality.

Ubuntu/Debian

sudo apt update
sudo apt install openjdk-21-jdk

macOS (using Homebrew)

brew install openjdk@21

Windows

Download and install from https://www.oracle.com/java/technologies/downloads/#java21 or use Chocolatey:

choco install openjdk21

Verify:

java -version

You should see something like:

openjdk version "21.0.x"
OpenJDK Runtime Environment ...
OpenJDK 64-Bit Server VM ...

Note: Java 17 or lower versions will not work as Pyserini requires Java 21.

Datasets

Built-in datasets are in autometrics/dataset/datasets (e.g., HelpSteer, SimpDA, ICLR, RealHumanEval, etc.). You can also construct your own via the Dataset class.

Minimal custom dataset example:

import pandas as pd
from autometrics.dataset.Dataset import Dataset

df = pd.DataFrame({
    'id': ['1', '2'],
    'input': ['prompt 1', 'prompt 2'],
    'output': ['response 1', 'response 2'],
    'reference': ['ref 1', 'ref 2'],
    'human_score': [4.5, 3.2]
})

dataset = Dataset(
    dataframe=df,
    target_columns=['human_score'],
    ignore_columns=['id'],
    metric_columns=[],
    name="MyCustomDataset",
    data_id_column="id",
    input_column="input",
    output_column="output",
    reference_columns=['reference'],
    task_description="Evaluate response quality",
)

Disk Caching

The library implements disk caching for all metrics to improve performance when running scripts multiple times. Key features:

• All metrics cache results by default in the ./autometrics_cache directory (configurable via AUTOMETRICS_CACHE_DIR)
• Cache keys include input/output/references and all initialization parameters
• Non-behavioral parameters are excluded automatically (name, description, cache config)
• You can exclude additional parameters via self.exclude_from_cache_key()
• Disable per-metric with use_cache=False
• Very fast metrics like BLEU/SARI may disable cache by default

See examples in autometrics/test/custom_metric_caching_example.py. For guidance on writing new metrics, see autometrics/metrics/README.md.

Where to Go Next

• Read the tutorial: examples/TUTORIAL.md (and examples/tutorial.ipynb)
• Browse built-in metrics under autometrics/metrics/
• Explore experiment drivers in analysis/

Citation

If you use this software, please cite it as below.

@software{Ryan_Autometrics_2025,
author = {Ryan, Michael J. and Zhang, Yanzhe and Salunkhe, Amol and Chu, Yi and Xu, Di and Yang, Diyi},
license = {MIT},
title = {{Autometrics}},
url = {https://github.com/XenonMolecule/autometrics},
version = {1.0.0},
year = {2025}
}