Skip to main content

Peer-based LLM cross-evaluation system

Project description

SlopRank

SlopRank is an evaluation framework for ranking LLMs using peer-based cross-evaluation and PageRank. It enables unbiased, dynamic, and scalable benchmarking of multiple models, fostering transparency and innovation in the development of AI systems.

You can use it with a large set of heterogeneous prompts to get overall rankings, or with smaller targeted sets to evaluate models for your specific use case.

Interactive Dashboard

Dashboard Preview

➡️ View Interactive Dashboard

Example Ranking:

=== PageRank Rankings ===
   model                                pagerank_score
0  o1-preview                           0.179404
1  gpt-4o                               0.178305
2  deepseek-chat                        0.167105
3  gemini-2.0-flash-thinking-exp-1219   0.164732
4  claude-3-5-sonnet-latest             0.155571
5  gemini-exp-1206                      0.154884

It supports pretty much all models, anything that can be run with the 'llm' library.

Features

  • Peer-Based Evaluation: Models evaluate each other's responses, mimicking a collaborative and competitive environment.
  • Customizable Scoring:
    • Numeric Ratings (1–10) for granular evaluation.
    • Upvote/Downvote for simple binary scoring.
  • Subset Evaluation: Reduce API costs by limiting the models each evaluator reviews.
  • Graph-Based Ranking: Endorsements are represented in a graph, and PageRank is used to compute relative rankings.
  • Scalable Benchmarking: Add more models or prompts with ease, maintaining flexibility and efficiency.
  • Graph Visualization: Visualize model endorsements with interactive and static graph visualizations.
  • Category-Based Analysis: Evaluate model performance across different prompt categories (reasoning, coding, etc.).
  • Statistical Confidence: Calculate confidence intervals and significance tests for model rankings.
  • Interactive Dashboard: Explore results through a web-based dashboard with interactive visualizations.

How It Works

  1. Prompt Collection: Define a set of questions or tasks to test the models.
  2. Model Responses: Each model generates a response to the prompts.
  3. Cross-Evaluation:
    • Each model evaluates the quality of other models' responses.
    • Evaluations are collected via predefined scoring methods.
  4. Graph Construction: Build a directed graph where nodes are models, and edges represent endorsements.
  5. Ranking: Apply the PageRank algorithm to rank models based on their relative endorsements.

Installation

Prerequisites

  • Python 3.8+
  • SimonW's llm library
  • networkx for graph computations
  • dotenv for environment variable management

Setup

SlopRank is on PyPI, so you can install it via:

pip install sloprank

From Source: If you prefer, clone this repo and install locally:

git clone https://github.com/strangeloopcanon/llmrank.git
cd sloprank
pip install .

API Keys Setup

Set up API keys using Simon Willison's llm tool. Example:

llm keys set anthropic 
llm keys set openai

Or create a .env file with:

OPENAI_API_KEY=your_openai_key
ANTHROPIC_API_KEY=your_anthropic_key

Usage

After installing, you can run the entire SlopRank workflow via the sloprank command. By default, SlopRank uses the models defined in DEFAULT_CONFIG. You can override this by passing --models with a comma-separated list.

Basic Usage

sloprank --prompts prompts.xlsx --output-dir results
  • --prompts prompts.xlsx tells SlopRank where to find your list of prompts.
  • --output-dir results puts all CSV and JSON outputs in the results/ folder.

If you want to override the default models:

sloprank --prompts prompts.xlsx --output-dir results \
         --models "chatgpt-4o,o1,claude-3-7-sonnet-latest, deepseek-reasoner, gemini-2.0-pro-exp-02-05" --visualize --confidence --categories

Configuration

  • Models: Update the MODEL_NAMES list to include the models you want to evaluate.
  • Prompts: Define your prompts in the raw_prompts list.
  • Evaluation Method: Choose between numeric ratings (EVALUATION_METHOD = 1) or upvotes/downvotes (EVALUATION_METHOD = 2).
  • Subset Evaluation: Toggle USE_SUBSET_EVALUATION to reduce evaluation costs.

Advanced Features

Visualization, Confidence Intervals, and Categories

Run SlopRank with all advanced features:

sloprank run --prompts prompts.xlsx --output-dir results --visualize --confidence --categories

Interactive Dashboard

Add the --dashboard flag to launch an interactive web dashboard:

sloprank run --prompts prompts.xlsx --output-dir results --dashboard

Launch the dashboard for existing results:

sloprank dashboard --output-dir results

Using Individual Tools

The examples/ directory contains standalone scripts for each advanced feature:

  1. Graph Visualization:

    python examples/generate_visualization.py
    
  2. Confidence Intervals:

    python examples/compute_confidence.py
    
  3. Prompt Categorization:

    python examples/prompt_categorization.py
    
  4. Dashboard Generation:

    python examples/generate_dashboard.py
    python examples/dashboard.py
    

Outputs

  • Ranked Models: A list of models ordered by their PageRank scores.
  • Graph Representation: A directed graph showing the flow of endorsements.
  • Processing Times: Benchmark of evaluation times for each model.
  • Interactive Visualizations: HTML-based interactive graphs with node and edge details.
  • Static Visualizations: PNG images of the endorsement graph.
  • Confidence Intervals: Statistical confidence bounds for model rankings.
  • Significance Tests: Statistical significance indicators between adjacent ranks.
  • Category Rankings: Model performance across different prompt categories.

Dashboard Details

The dashboard provides:

  • Overall model rankings with confidence intervals
  • Category-specific performance analysis
  • Interactive graph visualizations
  • Model comparison tools

Download Options

Applications

  • Benchmarking: Evaluate and rank new or existing LLMs.
  • Specialization Analysis: Test domain-specific capabilities (e.g., legal, medical).
  • Model Optimization: Identify strengths and weaknesses for targeted fine-tuning.
  • Public Leaderboards: Maintain transparency and foster healthy competition among models.

Development

Release Process

To build and release a new version of SlopRank to PyPI:

  1. Update the version number in pyproject.toml following semantic versioning
  2. Update the Changelog section below with all changes
  3. Clean previous builds: rm -rf build/ dist/ *.egg-info/
  4. Build the package: python -m build
  5. Validate the package: twine check dist/*
  6. Upload to PyPI: twine upload dist/*
  7. Create a GitHub release with the changelog info

Troubleshooting Releases

  • If you get permission errors during upload, check your PyPI credentials
  • If the build fails, ensure all dependencies are correctly listed in pyproject.toml
  • If the package fails validation, fix the issues before attempting to upload again

Version History

See the CHANGELOG.md file for a detailed version history and release notes.

Ideas for Contributions

Suggested Improvements

  1. Improve visualization options and customization.
  2. Add more statistical analysis methods.
  3. Develop a public leaderboard to showcase rankings.
  4. Enhance the web dashboard with more interactive features.
  5. Add support for multi-language evaluation by introducing localized prompts.
  6. Implement cost estimation and optimization features.

Contributions are welcome! If you have ideas for improving the framework, feel free to open an issue or submit a pull request.

Acknowledgments

Special thanks to:

  • SimonW for the llm library.
  • The AI community

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

sloprank-0.2.4.tar.gz (40.3 kB view details)

Uploaded Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

sloprank-0.2.4-py3-none-any.whl (36.5 kB view details)

Uploaded Python 3

File details

Details for the file sloprank-0.2.4.tar.gz.

File metadata

  • Download URL: sloprank-0.2.4.tar.gz
  • Upload date:
  • Size: 40.3 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/6.1.0 CPython/3.11.6

File hashes

Hashes for sloprank-0.2.4.tar.gz
Algorithm Hash digest
SHA256 73435937a796e66055b9ea764fd77c61c7bff036e075966b7e2937282751f0d4
MD5 aab3a2b79e405d61db7fbd2ee20f312b
BLAKE2b-256 3bbde61892103df2dfe21ad1e478bc2f016696efa10cac23e37451e7d9dad4a0

See more details on using hashes here.

File details

Details for the file sloprank-0.2.4-py3-none-any.whl.

File metadata

  • Download URL: sloprank-0.2.4-py3-none-any.whl
  • Upload date:
  • Size: 36.5 kB
  • Tags: Python 3
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/6.1.0 CPython/3.11.6

File hashes

Hashes for sloprank-0.2.4-py3-none-any.whl
Algorithm Hash digest
SHA256 098bf39dd282890744b503f378e58cd24640d5be12dde61cfdab014b08ccca22
MD5 328f07c88f4df35ab2a58b9cd21fd5f8
BLAKE2b-256 617ab4fa902e8abf8b76fe64ad8d1db193b56141304c93cbdfe89544c4e3acf5

See more details on using hashes here.

Supported by

AWS Cloud computing and Security Sponsor Datadog Monitoring Depot Continuous Integration Fastly CDN Google Download Analytics Pingdom Monitoring Sentry Error logging StatusPage Status page